# Surface Mobility Productivity Across the Mars Rover Fleet: What Drives Drive-Distance and Sols-per-Meter?

**A Dissertation Submitted to the COLLEGIUM Doctoral Board**

**Candidate:** JPL_AUTONOMY_EDL_03
**Program:** COLLEGIUM 1st Battalion
**NORTH STAR / JPL category:** Autonomous Systems and Robotics
**Methodological anchors (Hall of Shoulders):** Angrist and Pischke (design-based econometrics); Mokyr (economic history of technology)
**Stage:** Design-stage dissertation (no executed empirical results)
**Date:** 2026-06-15


## Abstract

A quarter-century of robotic service on the surface of Mars has been the work of many hands, and each mission has reached a little farther than the one before it. Each Mars surface mission has driven farther per sol than its predecessor, and the flight rovers (Sojourner, the two Mars Exploration Rovers, Curiosity, and Perseverance) advanced on two axes at once: their mechanical platforms grew larger and more capable, and their onboard autonomy software passed through several distinct generations. The standard narrative credits the rise in mobility productivity to better hardware, larger wheels, stronger actuators, a heavier chassis. This dissertation states and specifies a test of the rival claim that the dominant driver of cross-rover variation in mobility productivity is the autonomy-software generation and the quality of onboard hazard detection, not the mechanical platform. The contribution is a single falsifiable proposition: after conditioning on terrain class and on hardware covariates, autonomy-software generation explains the larger share of between-rover and within-rover variation in per-sol traverse productivity. The null is that productivity gains are attributable solely to mechanical platform improvement. The method is a two-way fixed-effects panel regression of per-sol traverse productivity on autonomy-generation indicators, hardware covariates, and terrain covariates, with rover and terrain-class fixed effects, identified primarily by a within-rover autonomous-drive-fraction contrast that holds the mechanical platform fixed at the level of the individual machine. Data are drawn from the Planetary Data System rover traverse and localization archives for the Mars Exploration Rovers, Curiosity, and Perseverance, the NASA Technical Reports Server performance reports for AutoNav and Enhanced Navigation, and NASA TechPort technology-readiness records. This is a design-stage dissertation: it specifies the estimator, the identification strategy, the bad-controls rule, the few-cluster inference, the pre-registered decision rule, and the threats to validity in full, and it presents expected results only as clearly labeled illustrations rather than as estimates produced from an assembled panel. The work matters because mission planners must know whether to buy mobility productivity with mass and power or with software, a decision with direct consequences for launch cost, mission risk, and the design of future rovers and the Mars Sample Return campaign, and one whose error is asymmetric: a software shortfall can sometimes be patched after landing, a hardware shortfall cannot.


## Table of Contents

- Abstract
- List of Tables and Figures
- Chapter 1: Introduction
  - 1.1 The chapter thesis
  - 1.2 The problem in full
  - 1.3 Institutional and historical context
  - 1.4 The research questions broken out explicitly
  - 1.5 The single falsifiable contribution stated as H0 and H1
  - 1.6 Significance for NASA, JPL, and the named stakeholders
  - 1.7 Scope and delimitations
  - 1.8 Definitions of key terms
  - 1.9 How the argument is built and what would defeat it
  - 1.10 The causal mechanism, named
  - 1.11 Confidence and the evidence that would move it
  - 1.12 Roadmap of the dissertation
- Chapter 2: Theoretical Framework
  - 2.1 The chapter's answer, stated first
  - 2.2 The problem this chapter must solve
  - 2.3 The Angrist-Pischke lens: from description to identification
  - 2.4 The Mokyr lens: autonomy as prescriptive knowledge resting on a propositional base
  - 2.5 The integrated conceptual model the empirical work will test
- Chapter 3: Literature Review
  - 3.1 The Mars Exploration Rover autonomy baseline
  - 3.2 The Mars Science Laboratory generation: inheritance, extension, and the terrain ceiling
  - 3.3 The Mars 2020 generation: moving computation into the drive
  - 3.4 Terramechanics and wheel-soil interaction: the source of the hardware and terrain covariates
  - 3.5 Surveys, path-planning generations, and the autonomy-generation taxonomy
  - 3.6 The Angrist-Pischke and Mokyr lenses applied to the domain literature
  - 3.7 Synthesis, the explicit gap, and the propositions that follow
- Chapter 4: Data and Measurement
  - 4.1 The PDS rover traverse and localization archives
  - 4.2 The NTRS AutoNav and Enhanced Navigation performance reports
  - 4.3 The TechPort mobility-technology TRL records
  - 4.4 Integrated operationalization of every variable
  - 4.5 Terramechanics-based construction of the hardware and terrain covariates
  - 4.6 Data quality, validation against known values, and access and ethics
  - 4.7 Construct validity and external corroboration of the productivity measure
- Chapter 5: Research Design and Identification
  - 5.1 The chapter thesis and the problem it answers
  - 5.2 The estimator and why it is chosen
  - 5.3 The specifications written out
  - 5.4 Identification assumptions argued formally
  - 5.5 Threats to validity and their mitigations
  - 5.6 The robustness battery
  - 5.7 Power and minimum-detectable-effect analysis
  - 5.8 The pre-registration commitment
  - 5.9 The computational and software plan
  - 5.10 Chapter synthesis: the design and its calibrated confidence
- Chapter 6: Analysis Plan and Expected Results
  - 6.1 Chapter thesis and what this chapter commits to in advance
  - 6.2 The step-by-step estimation procedure
  - 6.3 The fixed decision rule on the hypothesis
  - 6.4 Expected signs and the mechanism that generates them
  - 6.5 The three falsification checks, specified before estimation
  - 6.6 Power, the minimum detectable effect, and the limits of a small panel
  - 6.7 Illustrative simulation, descriptive generation profile and profile interpretation, and the unpopulated result tables
  - 6.8 Chapter synthesis and calibrated confidence
- Chapter 7: Discussion
  - 7.1 The chapter's answer, stated first
  - 7.2 Implications if H1 holds: software as the cheaper, retrofittable lever
  - 7.3 Implications if H0 holds: the mechanical lever, bounded and frozen
  - 7.4 Theoretical contribution back to each anchor framework
  - 7.5 Policy and mission implications for NASA, JPL, and stakeholders
  - 7.6 Full engagement with the rival explanations
  - 7.7 External-validity statement
  - 7.8 How this chapter advances the argument
- Chapter 8: Conclusion
  - 8.1 The chapter thesis
  - 8.2 Restating the contribution and what stands even if H1 is not confirmed
  - 8.3 Limitations, stated honestly
  - 8.4 A concrete future-research program
  - 8.5 Closing
- References
- Appendix A. Variable and Data Dictionary
- Appendix B. Derivations
- Appendix C. Instrument and Query Details
- Appendix D. Supplementary Tables


## List of Tables and Figures

This dissertation contains the following tables. It contains no figures; all design objects are presented as tables or as inline estimating-equation specifications.

| Table | Title | Location |
|---|---|---|
| 3.1 | MER (G1) generation evidence: capability, method, finding, limitation | Section 3.1 |
| 3.2 | MSL (G2) generation evidence: theme, method, finding, bearing on the gap | Section 3.2 |
| 3.3 | Mars 2020 (G3) generation evidence: theme, method, finding, bearing on the gap | Section 3.3 |
| 3.4 | Terramechanics evidence by function in the design | Section 3.4 |
| 3.5 | Survey and taxonomy evidence | Section 3.5 |
| 4.1 | Integrated operationalization of every variable | Section 4.4 |
| 6.1 | Nested decomposition of productivity variance (specified, unpopulated) | Section 6.7 |
| 6.2 | Within-rover autonomous-fraction estimates (specified, unpopulated) | Section 6.7 |
| 6.3 | Terrain-interaction profile (specified, unpopulated) | Section 6.7 |
| 6.4 | Falsification-check verdicts (specified, unpopulated) | Section 6.7 |
| A.1 | Variable and data dictionary | Appendix A |
| D.1 | Autonomy-generation crosswalk to TechPort TRL | Appendix D |

The four Chapter 6 result tables are specified but deliberately left unpopulated, consistent with the design-stage status of the dissertation. The tables in Chapter 3 are evidence-synthesis tables drawn from the published literature, not estimates from the panel.


# Chapter 1: Introduction

## 1.1 The chapter thesis

The robotic exploration of Mars is a long undertaking, sustained across decades by the patient effort of the institutions and people who design, build, and command these machines, and it deserves to be understood as carefully as it is carried out. Mobility is the binding constraint on a planetary rover's science return. The trend in Mars surface mobility productivity has been steeply upward across the flight fleet, and the unsettled question is not whether productivity rose but what caused it to rise. This dissertation argues that the durable share of that rise is driven principally by the autonomy-software generation and the quality of onboard hazard detection, not by the growth of the mechanical platform, and it states that argument as a single falsifiable proposition the design can reject. The stake is not academic. The two candidate causes, software and steel, are bought from different budgets, paid for at different points in the mission lifecycle, and recoverable to different degrees once a rover is on the surface, so the error of misattributing the cause is asymmetric and consequential for every rover NASA designs next. The chapter develops this thesis in order. It states the problem in full, locates it in the institutional and historical record of the Mars surface program, breaks out the research questions, fixes the falsifiable contribution as a null and an alternative, sets out the significance for NASA, the Jet Propulsion Laboratory (JPL), and the named stakeholders, bounds the scope and delimitations, defines the key terms, and closes with a roadmap of the dissertation. The register throughout is design-stage: no empirical estimates are reported in this work, and every expected value is labeled as an illustration of the design rather than as a finding produced from the assembled panel.

## 1.2 The problem in full

### 1.2.1 Mobility as the binding constraint

A rover that cannot reach a target cannot sample it, and a rover that reaches a target slowly has spent sols it cannot spend on science. Surface operations are organized around a daily planning cycle in which the ground team decides, before the next communication window, how far and where the rover will drive. The scarce quantity in that cycle is sols, and the planning literature treats onboard scheduling and decision-making as a way to relax that scarcity (Gao and Chien [\[50\]](#ref-gao2021)). Mobility productivity is therefore not one mission parameter among many. It is the currency in which the central activity of a surface mission, getting an instrument or a sampling tool onto the right piece of ground, is denominated. For the Mars 2020 Perseverance rover that central activity is caching: the rover must reach, characterize, and seal samples of Jezero crater material, and the count of sols consumed reaching each site is the count of sols not available for the sampling sequence itself (Farley et al. [\[45\]](#ref-farley2023)). The traverse simulation literature treats route length and traverse time as first-order mission-planning outputs precisely because they convert directly into the science schedule (Genova et al. [\[51\]](#ref-genova2013)). When mobility productivity rises, the science schedule loosens; when it stalls, the science schedule binds. That is the sense in which mobility is the binding constraint, and it is why the cause of the productivity rise is worth isolating rather than admiring.

### 1.2.2 The steep productivity trend across the fleet

The trend to be explained is steep and monotone. Sojourner, in 1997, traversed on the order of one hundred meters across its entire surface mission. The two Mars Exploration Rovers (MER), Spirit and Opportunity, drove tens of kilometers over operational lifetimes measured in years rather than the planned ninety sols (Crisp et al. [\[39\]](#ref-crisp2003); Squyres et al. [\[106\]](#ref-squyres2006)). Curiosity, the Mars Science Laboratory (MSL) rover, has driven more than thirty kilometers across Gale crater (Grotzinger et al. [\[55\]](#ref-grotzinger2012); Vasavada et al. [\[110\]](#ref-vasavada2014)). Perseverance has set single-sol and multi-sol traverse records that exceed anything earlier in the program, driving with a degree of onboard autonomy its predecessors did not possess (Farley et al. [\[45\]](#ref-farley2023); Verma et al. [\[112\]](#ref-verma2025)). Across these four missions, per-sol traverse productivity, the localized distance a rover converts from a sol of mission time, rose by a wide margin, and it rose at every generational step. The descriptive fact is not in dispute. What is in dispute is its decomposition.

### 1.2.3 The identification gap and its consequence

The problem this dissertation addresses can be framed as a gap between the current state of knowledge and the state a mission designer needs. The current state is a literature that describes each new rover as at once a better machine and a smarter one, and credits the combined improvement to the mission as a whole. The MER mission established a baseline that fused larger wheels and stronger actuators with onboard stereo hazard assessment and visual odometry (Crisp et al. [\[39\]](#ref-crisp2003); Maimone, Cheng, and Matthies [\[79\]](#ref-maimone2007)). Curiosity carried a much larger platform and an inherited, extended autonomy stack onto the harsher wheel-wear terrain of Gale crater (Grotzinger et al. [\[55\]](#ref-grotzinger2012); Arvidson et al. [\[18\]](#ref-arvidson2017)). Perseverance paired a refined chassis with Enhanced Navigation, the capability to process imagery and plan a path while the vehicle is still in motion (Verma et al. [\[112\]](#ref-verma2025)). Each of these descriptions is accurate, and each upgrades hardware and software at once. The desired state is different in kind: a designer allocating a fixed mass-and-power budget between heavier mechanical capability and more onboard computation needs to know which channel returns more productivity per unit of budget, holding terrain fixed. The gap between the two is an identification gap. The literature does not construct a comparison in which one channel varies while the other is held constant, because in the flight record the two channels never varied independently. The cost of leaving the gap open is a misallocation of the next rover's budget, and the misallocation is asymmetric: a software shortfall can sometimes be patched after landing through a flight-software update, whereas a hardware shortfall, an actuator too weak or a wheel too small, cannot be patched at all once the vehicle has left Earth. The problem, stated in full, is that the most decision-relevant question about twenty-five years of Mars surface mobility has an accurate narrative answer and no identified causal one.

## 1.3 Institutional and historical context

### 1.3.1 The Mars surface program as an evolving system
The Mars surface program is an unusual empirical setting because it is at once a sequence of distinct engineering artifacts and a continuous institutional lineage. Each flight rover was designed by overlapping teams at JPL, inherited code and lessons from its predecessor, and flew into an operational regime shaped by what the previous rover had learned. Sojourner proved that a wheeled vehicle could be commanded across Martian terrain at all. The MER mission scaled that proof into a two-rover, multi-year campaign and, in doing so, established the autonomy baseline that every later rover extended (Crisp et al. [\[39\]](#ref-crisp2003); Squyres et al. [\[106\]](#ref-squyres2006)). The MSL mission carried the lineage onto a far larger platform and into a science campaign of unusual duration and ambition (Grotzinger et al. [\[55\]](#ref-grotzinger2012); Vasavada et al. [\[110\]](#ref-vasavada2014)). The Mars 2020 mission added the most consequential autonomy step and tied surface mobility directly to the sample-caching objective of the broader Mars Sample Return campaign (Farley et al. [\[45\]](#ref-farley2023); Verma et al. [\[112\]](#ref-verma2025)). Because the lineage is continuous, the autonomy software of each generation is recognizably descended from the last, which is what makes a generation taxonomy meaningful; because the artifacts are distinct, each generation changed both the machine and the mind at once, which is what makes the causal decomposition hard. The institutional history is therefore not background color. It is the source of both the analyzable structure and the central identification problem of this dissertation.

### 1.3.2 The two competing narratives

Two explanations for the productivity rise circulate, usually together and rarely separated. The first is mechanical. Later rovers are larger, carry more capable actuators, use larger wheels, and have more available drive energy per sol; on this account, the rover simply became a better machine, and a better machine covers more ground. The terramechanics literature gives this narrative its rigor by modeling how commanded motion converts into actual progress through the wheel-soil interaction, and how that conversion degrades on soft and sloped ground (Ishigami et al. [\[61\]](#ref-ishigami2006) as cited in the prospectus; Gallina et al. [\[49\]](#ref-gallina2016)). The mechanical narrative is real: wheel wear and megaripple crossings throttled Curiosity's progress in ways that no amount of intelligence could fully overcome, and the terrain-interaction ceiling is a genuine physical constraint (Arvidson et al. [\[18\]](#ref-arvidson2017)). The second explanation is computational. Each rover generation introduced new autonomy software, from MER-class AutoNav with visual odometry, through the inherited and extended MSL stack, to Mars 2020 Enhanced Navigation, and on this account the rover became smarter, spending fewer sols per meter because it could plan around hazards in real time rather than waiting for a ground-in-the-loop command (Maimone, Cheng, and Matthies [\[79\]](#ref-maimone2007); Biesiadecki and Maimone [\[20\]](#ref-biesiadecki2007); Verma et al. [\[112\]](#ref-verma2025)). Both narratives are credible and both are partly right. The unresolved question is their relative weight. The literature's habit of crediting the combined improvement to the mission as a whole is precisely the description that forecloses the identification a designer needs. This dissertation takes the computational narrative as its alternative hypothesis and the mechanical narrative as its null, not because the mechanical channel is unreal but because the relative-weight question has never been posed as a falsifiable design.

## 1.4 The research questions broken out explicitly

The overarching research question asks which channel is more strongly associated with the steeply upward trend in Mars surface mobility productivity across the flight rover fleet: the mechanical platform, or the autonomy-software generation and onboard hazard-detection capability. After conditioning on terrain class and on hardware covariates, which channel accounts for the larger share of between-rover and within-rover variation in per-sol traverse productivity? The design makes a causal claim only where an estimator licenses it. The within-rover autonomous-fraction contrast holds hardware fixed at the level of the individual machine and is the sole identified comparison in the design. The between-rover generation decomposition is descriptive and associational; it reveals how productivity co-varies with generation but does not isolate the causal effect of generation, because no estimator in this design can separate the generation step from the simultaneous hardware upgrade. That overarching question decomposes into three operational questions, each corresponding to a distinct source of variation.

The first operational question is the between-rover decomposition, treated as descriptive throughout. Across the fleet, when per-sol mobility productivity is regressed on terrain fixed effects, then on terrain plus the hardware block, then on terrain plus hardware plus the autonomy-generation block, how much incremental explanatory share does the autonomy block add beyond what the hardware block already absorbs, and does adding the autonomy block attenuate the hardware coefficients? This question characterizes the associations among the channels in observational data and asks whether part of what looks mechanical in a naive comparison is correlated with the autonomy upgrade that accompanied every new rover. No claim of causal identification is attached to this decomposition; hardware and autonomy were upgraded together on every mission, and no between-rover estimator disentangles them.

The second operational question is the within-rover decomposition, and it is the primary one. Within a single rover, where the hardware is constant by construction, does the autonomous-drive fraction, the share of a drive's distance executed under onboard autonomous navigation rather than blind commanded motion, predict higher per-sol productivity after terrain class and terrain covariates are conditioned out? This question holds the machine fixed and asks whether the only thing that varies across autonomous and blind segments of the same rover's drives, namely how much work the software is doing, moves productivity.

The third operational question concerns the mechanism and its terrain signature. Is the autonomy effect, if present, largest in rougher and more hazard-dense terrain classes, where onboard hazard detection most reduces the penalty of blind driving, and smallest on benign flats, where blind commanded driving is already near-optimal? This question tests whether the proposed causal mechanism leaves the heterogeneity footprint it predicts, so that a confirmed effect is supported by its mechanism and not merely by a raw correlation.

These three questions are not independent tests of three claims. They are three views of one claim, chosen so that the strongest identifying variation, the within-rover autonomous-fraction contrast, carries the weight, while the between-rover decomposition supplies scope and the terrain-interaction signature supplies mechanism evidence. The design is built so that a single falsifiable proposition answers all three at once.

## 1.5 The single falsifiable contribution stated as H0 and H1

The contribution of this dissertation is one falsifiable proposition, stated as a pair of hypotheses for which the design specifies in advance what evidence would confirm or reject. The hypotheses are reproduced here in their operative form and are preserved verbatim across every chapter of the dissertation.

**H0 (null):** Conditional on terrain class, cross-rover and within-rover variation in per-sol mobility productivity is explained by mechanical platform covariates (wheel diameter, actuator and mass class, available drive energy) and not by autonomy-software generation. The autonomy-generation coefficients are jointly indistinguishable from zero once hardware covariates are included.

**H1 (contribution):** Conditional on terrain class and on the same hardware covariates, autonomy-software generation and onboard hazard-detection capability explain the larger share of the variation in per-sol mobility productivity. The autonomy-generation coefficients are jointly significant and remain so after hardware covariates are included, and they account for more of the between-rover and within-rover variance than the hardware block does.

The proposition is falsifiable in the strong sense, not merely the rhetorical one. The design states before any estimation the three pieces of evidence that would reject H1 and leave H0 standing. First, if the autonomous-drive-fraction coefficient is statistically indistinguishable from zero within rover, after terrain class and terrain covariates are conditioned out, the autonomy channel has no support where the machine is held fixed and H1 fails. Second, if the autonomy-generation coefficients lose joint significance the moment the hardware block enters the between-rover specification, the apparent autonomy effect is confounded with hardware and H1 fails. Third, if the terrain-interaction pattern is flat, with the autonomy effect no larger in hazard-dense terrain than on benign flats, the proposed mechanism is unsupported even if a raw correlation between autonomy generation and productivity exists. Naming these checks before estimation is the discipline that distinguishes a falsifiable contribution from an illustrated narrative, and it follows the practice that a credible empirical claim must name in advance the comparison that would prove it wrong (Angrist and Pischke [\[7\]](#ref-angrist2009)). The contribution is delivered as a complete research design rather than as executed estimates, and the document is explicit throughout that the expected results presented later are expectations stated to make the design concrete, not findings produced from the data.

A note on the form of the claim belongs here, because it governs how the rest of the dissertation reads. The claim is not that hardware does not matter. The mechanical channel is real, the terramechanics that bound it are real, and the null is constructed to give the mechanical channel every chance to account for the productivity rise on its own. The claim is comparative and conditional: after terrain and hardware are conditioned, the autonomy channel carries the larger share of the variation, and the within-rover variation, where hardware is held fixed at the level of the individual machine, shows an autonomy effect that survives. A comparative claim of this kind is rejectable on a clean criterion, the relative explanatory share and the survival of the within-rover coefficient, which is what makes it a contribution rather than an assertion.

## 1.6 Significance for NASA, JPL, and the named stakeholders

### 1.6.1 Why the answer changes the design lever

Mobility productivity is a currency that mission planners spend, and the question of which channel produces it is a question about which lever a designer should pull. If productivity is bought with mass and power, the design lever is mechanical, and the cost is paid at launch, where every kilogram is dear and every design choice is frozen at the moment the vehicle leaves Earth. If productivity is bought with software, the design lever is computational, the cost is paid in development and verification rather than at launch, and the capability can in principle be delivered to a rover already on the surface through a flight-software update, as the Mars program has done repeatedly. The Enhanced Navigation capability on Perseverance is the proof of concept that the computational lever is real and deliverable, raising the autonomous-drive fraction and the achievable distance per sol relative to earlier generations (Verma et al. [\[112\]](#ref-verma2025)). Knowing which lever dominates is therefore a precondition for rational investment in future surface systems, not a refinement of an already-settled allocation. The significance of the contribution is that it converts a narrative crediting each rover's success to the whole machine into a design that asks where the next increment of productivity should be bought.

### 1.6.2 The Mars Sample Return stake

The most immediate stakeholder consequence is the Mars Sample Return campaign, whose caching activity is conducted by Perseverance and whose currency is sols. Every sol that mobility productivity frees is a sol available for reaching, characterizing, and sealing a sample, the central activity on which the entire campaign depends (Farley et al. [\[45\]](#ref-farley2023)). A rover that needs fewer sols per meter has more sols for sampling, instrument placement, and caching. If the autonomy channel dominates, then the productivity payoff of onboard autonomy is not only the meters it adds on a given drive but the planning-cycle time it frees, because a rover trusted to handle hazards autonomously needs less conservative, more sparsely specified commands, which loosens the daily planning bottleneck that the operations literature identifies as the binding scarcity (Gao and Chien [\[50\]](#ref-gao2021)). That second-order effect, the planning effort consumed per meter, is not captured by the simple meters-per-sol measure this design uses, and the dissertation is explicit that a confirmed autonomy effect would, if anything, understate the full operational value of the autonomy channel.

### 1.6.3 The asymmetry of the policy error

The significance of getting the answer right is sharpened by the asymmetry of the error. If a mission planner wrongly concludes that productivity is mechanical when it is in fact computational, the planner over-invests in mass and power, pays the cost at launch, and forgoes the cheaper software lever and its retrofit option. If a planner wrongly concludes that productivity is computational when it is in fact mechanical, the planner under-builds the platform and discovers on the surface that no software update can compensate for an actuator that is too weak or a wheel that is too small. These two errors are not symmetric in their recoverability. A software shortfall can sometimes be patched after landing; a hardware shortfall cannot. This asymmetry is itself a reason to estimate the two channels carefully rather than rely on the combined narrative, and it is the reason the contribution is decision-relevant rather than merely descriptive. The named stakeholders, the rover and surface-systems designers at JPL, the Mars Sample Return campaign planners, and the program managers who allocate mass and power budgets before launch, are precisely the actors for whom this asymmetry is operative.

### 1.6.4 The historical-economics significance

Beyond the immediate mission decision, the contribution speaks to a question in the economic history of technology, and the dissertation reads its result through that frame. The distinction between propositional knowledge, the understanding of why something works, and prescriptive knowledge, the technique that does the work, supplies an interpretation of why the two channels should differ in their returns (Mokyr [\[85\]](#ref-mokyr2002)). Mechanical platform improvement is prescriptive refinement of a known kind: a larger wheel, a stronger actuator, a heavier chassis are refinements of an established technique whose returns are real but bounded by the same physics that bounded the previous design. Autonomy-software generation, by contrast, rests on a deepening propositional base in perception, planning, and terrain modeling, and is therefore extensible in a way that a wheel is not. The same flight processor that runs Enhanced Navigation can in principle run a still better planner uploaded after landing, whereas a wheel cannot be made larger after launch. If H1 holds, the productivity history of the Mars fleet becomes an instance of a general pattern in which the durable gains come from the extensible, knowledge-intensive layer rather than the bounded mechanical one. That reading carries beyond the rover fleet because it offers a structured conjecture, to be tested rather than assumed, about whether the autonomy channel should transfer better than the hardware channel to lunar rovers, to future Mars rovers, and to terrestrial off-road autonomy.

## 1.7 Scope and delimitations
The scope of this dissertation is deliberately bounded, and the delimitations are stated plainly because they govern what the eventual estimates could and could not claim. The empirical setting is the flight rover fleet on the surface of Mars: Sojourner, the two Mars Exploration Rovers, Curiosity, and Perseverance. Three of these are full panel members: the MER pair, Curiosity, and Perseverance, whose traverse archives in the Planetary Data System support per-drive-sol construction of the dependent variable. Sojourner is carried as a boundary case rather than a full panel member, because its surface mission predates the standardized traverse-product archive and its drive-level record is too thin to support the same panel construction; it is treated qualitatively and is not promised as a fourth full panel member. The unit of analysis is the drive-sol, one rover on one sol with its commanded drives aggregated to the sol, and the panel is an unbalanced panel of drive-sols nested within rovers. The individual drive is a secondary unit used for robustness where the archive resolves multiple drives within a sol.

Three delimitations bound external validity and are acknowledged at the outset. First, the cross-section of rovers is small. There are four rovers and three full panel members, so between-rover identification rests on few clusters and leans heavily on within-rover variation; the inference procedure is chosen accordingly. Second, the fleet is Mars-specific and site-specific. The rovers drove three landing sites on one planet, and generalization to lunar rovers, to future Mars rovers with different processors, or to terrestrial autonomy is not warranted without re-estimation. The historical-economics frame offers a conjecture that the autonomy channel should generalize better than the hardware channel, but that conjecture is a hypothesis for future work, not a result this dissertation delivers. Third, the productivity construct is partial. Meters-per-sol and its inverse are defensible and directly available from the archives, but they ignore the scientific value of where the rover went and the ground-team planning effort consumed per meter; the dissertation uses the simpler distance-and-sols measure because it is directly available and is not contaminated by the analyst's model of ground-team behavior, and it flags the richer construct as future work. Beyond these, the dissertation does not model the EDL phase that delivers each rover to the surface, does not estimate the scientific yield of the traverses, and does not propose or evaluate any new rover or navigation system. It is an econometric research design over existing public archives, and it confines itself to that.

## 1.8 Definitions of key terms

The following definitions are fixed across the dissertation and are condensed here from the measurement specification developed in full in Chapter 4.

**Mobility productivity (the dependent variable).** Operationalized two ways, estimated in parallel: meters traversed per sol, the rate at which the rover converts mission time into distance; and sols per meter, its inverse, the quantity a planner budgets. Distance is the localized path length from the Planetary Data System traverse product, the actual path driven, not the straight-line displacement.

**Autonomy-software generation (the treatment).** A categorical indicator with three levels: G1, MER-class AutoNav with visual odometry; G2, the MSL inherited and extended stack; and G3, the Mars 2020 Enhanced Navigation system. A continuous secondary measure, the autonomous-drive fraction, is the share of a drive's distance executed under onboard autonomous navigation rather than blind commanded motion.

**Hardware covariates.** The block of mechanical attributes that the design conditions on: wheel diameter, mass class, actuator class, and nominal available drive energy per sol.

**Terrain covariates and the terrain fixed effect.** Terrain class is the dimension absorbed by the fixed effect; the continuous terrain covariates are slope and a physical-properties index, drawn from orbital basemaps and archived terrain characterizations.

**Mediator, not control.** Realized slip and realized wheel-soil interaction are post-treatment mediators. They are recorded but stay on the outcome side of the analysis and never enter the right-hand side of the regression, in keeping with the bad-controls discipline that forbids conditioning on a variable that is itself a consequence of the treatment (Angrist and Pischke [\[7\]](#ref-angrist2009)).

**Drive-sol.** The primary unit of analysis: one rover, one sol, with its one or more commanded drives aggregated to the sol.

**Blind commanded motion versus autonomous navigation.** Blind commanded motion is a drive segment executed under a ground-specified path without onboard hazard re-planning; autonomous navigation is a segment in which the rover detects hazards and selects or adjusts its path onboard. The distinction is the substance of the autonomous-drive fraction and is the operational meaning of the autonomy treatment (Biesiadecki and Maimone [\[20\]](#ref-biesiadecki2007); Verma et al. [\[112\]](#ref-verma2025)).

## 1.9 How the argument is built and what would defeat it

The central argument of this dissertation can be set out plainly, together with the conditions under which it would fail. The argument is that the autonomy-software generation, not the mechanical platform, drives the durable share of the productivity rise. It rests on the convergent record that productivity rose at every generational step while a distinct, recognizable autonomy capability was introduced at each step, from visual odometry and stereo hazard assessment on MER (Maimone, Cheng, and Matthies [\[79\]](#ref-maimone2007)), through the inherited MSL stack (Grotzinger et al. [\[55\]](#ref-grotzinger2012); Vasavada et al. [\[110\]](#ref-vasavada2014)), to Enhanced Navigation on Perseverance (Verma et al. [\[112\]](#ref-verma2025)). What licenses the move from that record to a causal claim is the design-based principle that a credible attribution requires a comparison holding the rival cause fixed, supplied here by the within-rover autonomous-fraction contrast in which hardware is constant at the level of the individual machine (Angrist and Pischke [\[7\]](#ref-angrist2009)). That principle is anchored in the broader credibility-revolution practice of naming the counterfactual, forbidding bad controls, and using inference appropriate to the cluster structure. The argument is qualified, and the qualification is protected throughout: the claim holds conditional on terrain class and hardware covariates, it is comparative rather than absolute, and the supporting evidence is design-stage, so the confidence attached to it is the confidence of a well-specified design, not of an executed estimate. The conditions that would defeat it are stated as the three falsification checks of Section 1.5, and any one of them, if realized in the data, moves the verdict toward H0.

The chapters that follow build the case in stages. They first establish that the productivity rise is genuine and that the autonomy and hardware channels have been described together and never separated (Crisp et al. [\[39\]](#ref-crisp2003); Gao and Chien [\[50\]](#ref-gao2021); Verma et al. [\[112\]](#ref-verma2025)), and that separating them matters because sols are the binding scarcity for Mars Sample Return caching and instrument placement, which makes the productivity decomposition a decision the next mission must act on (Farley et al. [\[45\]](#ref-farley2023); Genova et al. [\[51\]](#ref-genova2013)). They then show how the design reaches the causal mechanism: the within-rover autonomous-fraction contrast holds hardware fixed and isolates the software channel, while the bad-controls rule keeps realized slip on the outcome side (Angrist and Pischke [\[7\]](#ref-angrist2009); Maimone, Cheng, and Matthies [\[79\]](#ref-maimone2007); Helmick et al. [\[59\]](#ref-helmick2004)), and why that contrast improves on the alternatives, dominating both the combined-narrative description and the naive between-rover regression that cannot separate collinear channels. What risk remains is made bounded and visible through few-cluster inference and an explicit threat-to-validity treatment rather than left implicit. Each subsequent chapter carries one stage of this argument forward, and the conclusion draws them together.

## 1.10 The causal mechanism, named

A causal claim must name a mechanism, not rest on a bare correlation, and the mechanism here is specific and testable. The driver is the advance in autonomy-software generation from G1 to G2 to G3, culminating in Enhanced Navigation that processes imagery and plans a path while the rover is still in motion (Verma et al. [\[112\]](#ref-verma2025)). The mechanism is that a larger share of each drive is executed under onboard hazard detection and real-time path planning rather than blind commanded motion, so that the rover no longer has to stop, image, think, and only then move. The observable effect is a higher autonomous-drive fraction and more localized meters per sol, concentrated in hazard-dense terrain classes where the blind-driving penalty is largest. The operational consequence is fewer sols consumed per meter, which frees sols for sampling, instrument placement, and caching in the daily planning cycle. The strategic implication is that surface mobility productivity is bought more cheaply with software than with mass, the cost is paid in development rather than at launch, and the capability can be uploaded to a rover already on the surface. Each link in this chain is an empirical commitment the design can examine: the autonomous-fraction-to-productivity link is the within-rover test, the terrain concentration is the heterogeneity test, and the generational pattern is the between-rover decomposition. Where the design can establish only correlation rather than mechanism, as in the coarse between-rover generation contrast where hardware and autonomy are collinear by construction, the dissertation says so and downgrades its confidence accordingly, which is why the within-rover contrast and its terrain signature, not the between-rover correlation, carry the causal weight.

## 1.11 Confidence and the evidence that would move it

The confidence this dissertation attaches to its central proposition is, at the design stage, moderate, and it is calibrated to the grade of evidence a design can offer rather than to the grade an executed estimate would offer. The proposition rests on three supports of differing strength. The strongest is the within-rover identification logic, which is high-confidence as a design because it holds hardware fixed at the level of the individual machine and reduces to a comparison the credibility-revolution literature endorses (Angrist and Pischke [\[7\]](#ref-angrist2009)). The second is the convergent domain record that the autonomy generations are genuinely distinct capabilities and not relabelings, which is well-supported by the engineering literature and by the independent technology-readiness records the design uses as a robustness check (Maimone, Cheng, and Matthies [\[79\]](#ref-maimone2007); Biesiadecki and Maimone [\[20\]](#ref-biesiadecki2007); Verma et al. [\[112\]](#ref-verma2025)). The weakest support, and the one that bounds overall confidence, is the availability and quality of a clean per-drive autonomous-fraction series across all generations; the public performance reports document autonomous-drive capability and behavior, but a uniform per-drive fraction series across every generation is not guaranteed in the public record, and the design carries a measurement-error caveat accordingly. What would raise confidence is the assembly of that fraction series with low measurement error and the realization of the predicted terrain-interaction signature in the data. What would lower it is a null within-rover coefficient, a collapse of the autonomy block when hardware enters, or a flat terrain-interaction pattern, each of which is a pre-specified rejection condition. Stating in advance both the confidence and the evidence that would move it is part of the design discipline this dissertation adopts, and it is the reason the work is presented as a falsifiable design rather than as a confirmed result.

## 1.12 Roadmap of the dissertation

The dissertation proceeds in eight chapters and a backmatter, organized so that the design is built before any expected result is discussed and so that interpretation follows estimation rather than preceding it.

Chapter 2 develops the two theoretical anchors, and does so because they do non-redundant work. The design-based econometrics anchor tells us how to estimate the relative contribution of the two channels without fooling ourselves, through a named counterfactual, a within-unit comparison, a strict bad-controls rule, and few-cluster-aware inference. The economic-history-of-technology anchor tells us how to interpret the estimate once we have it, through the distinction between bounded prescriptive refinement of mechanical hardware and extensible improvement resting on a deepening propositional base in perception and planning. Chapter 2 integrates the two into the conceptual model the empirical work will test.

Chapter 3, the literature review and the longest chapter, surveys the domain literature thematically by autonomy generation, then across the cross-cutting terramechanics and terrain theme, then through the surveys that establish the autonomy-generation taxonomy, and it closes with the explicit gap statement that motivates the propositions: the literature is rich in description and thin in identification.

Chapter 4 sets out the data and measurement, dataset by dataset, covering the Planetary Data System traverse and localization archives that supply the dependent variable, the NASA Technical Reports Server performance reports that supply the autonomy-generation classification and the autonomous-drive fraction, and the NASA TechPort technology-readiness records that supply the independent maturity classification used as a robustness check. It presents the full variable-operationalization table and the data-quality and validation procedures.

Chapter 5 specifies the research design and identification in full: the two-way fixed-effects estimator, the within-rover and between-rover specifications, the formal identification assumptions, the collinearity problem and its two-pronged resolution, the threats to internal, external, construct, and statistical-conclusion validity, the robustness battery, and the power and pre-registration commitments.

Chapter 6 presents the analysis plan and the expected results. It gives the step-by-step estimation procedure, fixes the decision rule on H1 against H0, sets out the expected signs with their mechanism reasoning, and specifies the three falsification checks, with every number labeled as an illustration of the design rather than as an executed estimate and every result table specified but left unpopulated by design.
Chapter 7 discusses the implications under both possible outcomes, reads each outcome through both anchor frameworks, draws out the policy and mission implications for NASA, JPL, and the named stakeholders, engages fully with the rival explanations of inseparability, terrain selection, and ground-team operational learning, and states the bounded external-validity conjecture.

Chapter 8 concludes by restating the contribution and what stands regardless of the eventual verdict, stating the limitations honestly, and laying out a concrete future-research program that runs from executing the design on the full panel through the richer planning-effort-per-meter productivity construct to lunar and terrestrial replication.

The backmatter compiles the full reference list from the project corpus with clickable identifiers and provides the data dictionary, the identification and inference derivations, the archive access details, and the autonomy-generation taxonomy crosswalk.

The dissertation has a single through-line. Mars surface mobility productivity rose steeply and monotonically; the cause is contested between steel and software; the contest has an accurate narrative answer and no identified causal one; and this work supplies the design that can produce the identified answer and can be rejected if it is wrong. The contribution is the design and the falsifiable proposition it tests, not a coefficient, and the chapters that follow build that design in the order a careful reader would want it built.



# Chapter 2: Theoretical Framework

## 2.1 The chapter's answer, stated first

This dissertation rests on two theoretical anchors, and the claim of this chapter is that they do different, non-substitutable work. Angrist and Pischke supply the discipline of estimation: they tell us how to recover the relative contribution of two entangled channels, autonomy software and mechanical hardware, without fooling ourselves into crediting one for what the other did. Mokyr supplies the discipline of interpretation: he tells us what a recovered estimate would mean for the future trajectory of the technology, why the two channels should be expected to differ in the durability of their returns, and why an advantage for the software channel would be an instance of a recurring historical pattern rather than a one-off fact about four rovers. The chapter's thesis is therefore not that either framework is correct in the abstract, but that the empirical question posed by this dissertation cannot be answered with only one of them. A purely statistical decomposition that names a winner is inert without an account of why the channels should differ. A purely historical claim that software is the durable lever is merely suggestive without a credible estimate that holds terrain and hardware fixed. The conceptual model the empirical work will test is the join of the two: a within-rover, terrain-conditioned comparison that isolates the software channel (the Angrist-Pischke contribution), interpreted through the propositional-versus-prescriptive distinction that predicts which channel's gains should persist and transfer (the Mokyr contribution).

The chapter develops this thesis in three movements. Section 2.3 develops the Angrist-Pischke lens in depth: its primary sources, the four disciplines it imposes (named counterfactual, bad-controls avoidance, few-cluster-aware inference, and the credibility standard that a claim must specify in advance what would refute it), and exactly how each transfers to a panel of three flight rovers. Section 2.4 develops the Mokyr lens: the propositional-prescriptive distinction, the argument that a wheel is a finished technique with bounded returns while autonomy rests on a widening propositional base and is therefore extensible and retrofittable, and the access-cost condition under which such gains persist. Section 2.5 builds the integrated conceptual model that the rest of the dissertation tests, mapping each lens onto a specific component of the estimation design and stating the model's testable implications. The register throughout is design-stage: no estimate in this chapter is executed, and every quantitative expectation is labeled as illustrative.

## 2.2 The problem this chapter must solve

Stating the problem in the explicit current-state, desired-state, gap, consequence form makes the choice of two anchors rather than one follow directly from it.

The current state is that the space-robotics literature describes each Mars rover as simultaneously a better machine and a smarter one, and credits the combined improvement to the mission as a whole. Crisp et al. [\[39\]](#ref-crisp2003) describe the Mars Exploration Rover mission as an integrated leap in platform and capability; Grotzinger et al. [\[55\]](#ref-grotzinger2012) do the same for Curiosity; Verma et al. [\[112\]](#ref-verma2025) describe Enhanced Navigation on Perseverance as a capability advance riding on a larger, more capable rover. The reviews that take the long view, notably Gao and Chien [\[50\]](#ref-gao2021), narrate a trajectory in which platform and autonomy advance together. Each of these accounts is accurate. None of them is an identification. The desired state is a framework that can do two things at once: estimate how much of the productivity gain is attributable to the autonomy channel alone after holding terrain and hardware fixed, and interpret that estimate in a way that tells a mission designer what it implies for where to spend the next increment of mass, power, and development effort.

The gap is that no single body of theory available to this dissertation does both. Design-based econometrics is built to recover causal contributions from observational variation, but it is silent on why two technical channels should differ in the durability or transferability of their returns; it would estimate a coefficient and stop. The economic history of technology is built to explain why some technical improvements compound and others plateau, but it offers no estimator and no rule for separating collinear channels from archival data; it would interpret a fact it could not itself establish. The consequence of trying to proceed with only one lens is concrete. With only Angrist and Pischke, a confirmed software advantage would be a number without a meaning, and a mission planner would have no principled reason to expect the advantage to recur in the next rover or transfer to a lunar platform. With only Mokyr, a historical narrative favoring software would rest on the very combined-narrative description the dissertation set out to improve upon, and would be vulnerable to the objection that the analyst simply chose which factor to privilege. The two anchors are therefore not a stylistic pairing but a logical necessity: the estimation lens supplies a result the interpretation lens cannot establish, and the interpretation lens supplies a meaning the estimation lens cannot generate.

## 2.3 The Angrist-Pischke lens: from description to identification

### 2.3.1 The framework and its primary sources

The claim of this section is that the design-based, or credibility-revolution, school of empirical economics supplies exactly the estimation discipline this dissertation needs, and that its discipline transfers to a rover panel without distortion because the panel's central problem, two collinear channels upgraded together, is the same problem the school was built to solve in labor and program evaluation.

The primary sources form a coherent body. Imbens and Angrist [\[60\]](#ref-imbensangrist1994) established the local-average-treatment-effect framework, formalizing what a causal estimate identifies when treatment is not randomly assigned but is driven by an instrument: the effect for the subpopulation whose treatment status the instrument moves. Angrist, Imbens, and Rubin [\[9\]](#ref-angristimbensrubin1996) extended this into the potential-outcomes account of instrumental variables, making explicit the assumptions, exclusion, monotonicity, and a nonzero first stage, under which an observational comparison recovers a causal quantity rather than a contaminated correlation. The two textbook treatments, Mostly Harmless Econometrics [\[7\]](#ref-angrist2009) and Mastering 'Metrics [\[8\]](#ref-angrist2014), distilled these foundations into a working discipline for applied researchers, and the manifesto, The Credibility Revolution in Empirical Economics [\[10\]](#ref-angristpischke2010), stated the school's central methodological commitment: that the credibility of an empirical claim rests on the transparency and defensibility of its research design, not on the sophistication of its estimator or the size of its dataset.

The interpretation this dissertation places on that body of work is that it converts a vague aspiration, to learn what caused something, into a small set of enforceable rules. The school's enduring contribution is less any single estimator than the insistence that every causal claim be accompanied by an explicit account of the comparison that identifies it and the assumptions under which that comparison is valid. That insistence is what the combined-narrative rover literature lacks, and it is what this section imports.

### 2.3.2 Discipline one: name the counterfactual and the comparison

The first discipline is that every causal claim must name its counterfactual and the comparison that identifies it [\[7\]](#ref-angrist2009). Treating this as binding follows from a simple observation: without a named comparison, a claim that autonomy drove productivity is indistinguishable from a claim that hardware did, because both channels moved in the same direction at the same time. The principle that connects this requirement to credible inference is the potential-outcomes logic of Imbens and Angrist [\[60\]](#ref-imbensangrist1994), [\[9\]](#ref-angristimbensrubin1996): a causal effect is defined as a contrast between potential outcomes under treatment and under control for the same units, and an estimate is credible only to the extent that the data approximate that contrast. The accumulated record of the credibility revolution bears this out, in which the move from associational regression to designs with named counterfactuals demonstrably changed which empirical claims survived scrutiny [\[10\]](#ref-angristpischke2010).

Transferred to the rover panel, the discipline specifies the comparison precisely. The counterfactual of interest is the per-sol mobility productivity a given rover would have achieved on a given terrain class under a different autonomy-software generation, holding its mechanical platform fixed. The comparison that approximates this counterfactual is variation in per-sol productivity across drive-sols that share a terrain class but differ in how much of the drive the software executed autonomously. Within a single rover, the mechanical platform is constant by construction, so a comparison of autonomous-heavy and blind-commanded segments on the same machine in the same terrain class is the closest available approximation to the named counterfactual. One qualification is essential and is stated here rather than buried: the comparison is an approximation, not a randomized contrast, because the share of a drive executed autonomously is chosen by the ground team rather than assigned at random, and that choice is correlated with terrain. The objection this discipline forces us to anticipate is that the autonomous fraction is endogenous; the design's answer, developed in Chapter 5, is to condition out the terrain-driven component of that choice through terrain-class fixed effects and terrain covariates so that the residual variation in autonomous fraction is closer to as-good-as-random with respect to productivity. Confidence in this transfer is moderate at the design stage: the comparison is well specified and the endogeneity is named and partially addressable, but whether the residual variation is in fact close to random is itself an assumption the executed analysis must defend, not a settled fact.

### 2.3.3 Discipline two: do not condition on bad controls

The second discipline is the bad-controls rule: never condition on a variable that is itself a consequence of the treatment [\[7\]](#ref-angrist2009). This rule decides, in advance, which rover-panel variables may enter the right-hand side and which must stay on the outcome side, and violating it would bias the autonomy estimate toward the null in a way that would manufacture a spurious confirmation of H0. The formal result, stated plainly in Mostly Harmless Econometrics, is that controlling for a post-treatment variable opens a non-causal path between treatment and outcome and absorbs part of the very effect under study, which follows directly from the potential-outcomes definition of a covariate as admissible only if it is fixed prior to treatment assignment. The wide empirical record of attenuation produced by over-controlling in program evaluation, which the school documents as a recurring error, confirms how often this trap is sprung.

The transfer to the rover panel is direct and consequential. Realized slip and realized wheel-soil interaction are produced by the drive: the autonomy software, by choosing how and where to drive, partly determines how much the wheels slip. Conditioning on realized slip would therefore absorb part of the autonomy effect and bias the estimate toward zero, which in this design is the direction that falsely supports the mechanical-only null. The mechanism of the bias is specific. The driver is the bad control entering the regression; the mechanism is that the control sits on the causal path from autonomy choice to productivity, so the regression attributes to the control what belongs to the treatment; the observable effect is an attenuated autonomy coefficient; the operational consequence is a false rejection of H1; the strategic implication is that a mission designer would wrongly conclude that productivity is mechanical and over-invest in mass. Because the consequence is this serious, the design draws a hard line: a priori terrain class and commanded-drive parameters are eligible controls because they are fixed before the drive executes, while anything realized during or after the drive, slip, drive duration, wheel wear, is a mediator that stays on the outcome side. This line is not a refinement of the design; it is the design's spine, and the bad-controls discipline is what justifies drawing it where it is drawn. Confidence that the line is correctly placed is high, because the distinction between a priori and realized quantities is observable in the data archive and the bad-controls result is a theorem, not an empirical regularity.

### 2.3.4 Discipline three: be honest about few-cluster inference
The third discipline concerns inference rather than identification. With repeated drives nested within a very small number of rovers, conventional standard errors understate uncertainty, and honest inference requires methods built for few clusters together with an explicit caveat. The structure of the panel forces this. There are three full panel members and a boundary case, so the effective number of independent clusters for between-rover contrasts is on the order of three, far below the asymptotic regime that conventional cluster-robust standard errors assume. The relevant econometric theory concerns estimation and inference when treatment timing and treatment effects are heterogeneous and clusters are few. De Chaisemartin and D'Haultfoeuille [\[29\]](#ref-chaisemartin2020) show that two-way fixed-effects estimators can be badly behaved under heterogeneous treatment effects, computing negative weights on some comparisons; Goodman-Bacon [\[54\]](#ref-goodmanbacon2021) decomposes the two-way fixed-effects estimand into a weighted average of two-group comparisons and shows which comparisons drive it; Callaway and Sant'Anna [\[23\]](#ref-callawaysantanna2021) supply estimators for settings with multiple periods and staggered adoption that avoid the contaminating comparisons. These results have reshaped applied practice in difference-in-differences settings, which are structurally analogous to a staggered rollout of autonomy generations across the fleet.

The transfer is bounded, and it is the point at which this dissertation must avoid overreach. The rover panel is not a textbook staggered difference-in-differences design; it is a small panel with a categorical generation treatment that is collinear with rover identity. The heterogeneous-effects literature transfers in two ways. First, it warns that a naive two-way fixed-effects regression of productivity on generation could place perverse weights on cross-rover comparisons, so the design must inspect the implied weights rather than trust the headline coefficient [\[29\]](#ref-chaisemartin2020), [\[54\]](#ref-goodmanbacon2021). Second, it motivates the within-rover autonomous-fraction contrast as the cleaner identifying variation precisely because it sidesteps the cross-rover weighting problem: within a single machine there is no staggered-adoption contamination, only variation in how much the software did. The few-cluster problem proper is handled by wild-cluster bootstrap inference for the between-rover contrasts, with the few-cluster caveat reported rather than hidden. The qualifier the school demands, and that this dissertation accepts, is that no inferential method recovers precision the data do not contain. With three clusters, the between-rover contrast will be imprecise, and the design's honest posture is to lean on the within-rover variation for its primary test and to treat the between-rover generation contrast as a complement reported with wide, few-cluster-aware uncertainty. Confidence in this transfer is high as to the direction of the problem and the appropriateness of the remedies, and deliberately modest as to how much precision any remedy can deliver from three rovers.

### 2.3.5 Discipline four: specify in advance what would refute the claim

The fourth discipline is the credibility standard itself: a claim is credible to the degree that its design specifies, before estimation, what evidence would reject it [\[7\]](#ref-angrist2009), [\[10\]](#ref-angristpischke2010). This standard is what makes the dissertation's contribution falsifiable rather than merely illustrated, and it is the discipline most often missing from the combined-narrative rover literature, which describes outcomes without naming the comparison that could prove its causal attributions wrong. The credibility-revolution literature argues that pre-specification disciplines the analyst against the temptation to fit a story to whatever the data show, on the principle that a hypothesis which cannot be refuted by any conceivable result carries no information, while one that names its own defeat conditions in advance carries the maximum the data can supply. The school's documented shift toward pre-analysis plans and design transparency as the markers of credible work confirms how far this standard now reaches.

Transferred to this dissertation, the discipline produces the three falsification checks that are stated, by design, before any estimation: if the within-rover autonomous-fraction coefficient is indistinguishable from zero after terrain conditioning, the autonomy channel has no support and H1 fails; if the between-rover autonomy-generation block loses joint significance the moment hardware covariates enter, the apparent autonomy effect is confounded with hardware and H1 fails; and if the terrain-interaction pattern is flat rather than concentrated in hazard-dense terrain classes, the proposed mechanism is unsupported even where a raw correlation exists. Naming these checks in advance is the credibility standard doing its work. Pre-specification protects against overfitting but not against a genuinely underpowered test: a null within-rover coefficient could reflect either a real absence of an autonomy effect or a panel too small to detect one, and the design must report which, through the minimum-detectable-effect discussion deferred to Chapter 5. Confidence that the falsification checks are well specified is high; confidence that the eventual data will be powered to make them decisive is, honestly, only moderate, and the dissertation says so.

### 2.3.6 What the lens does and does not provide

This section advances the third of the dissertation's commitments: that the design addresses the causal mechanism rather than merely correlating with it. The Angrist-Pischke lens supports that commitment by supplying the within-rover comparison that holds hardware fixed, the bad-controls rule that keeps realized slip off the right-hand side, and the inferential honesty that prevents three rovers from being mistaken for a large sample. What the lens does not provide, and what no estimator can provide, is a reason to expect the recovered estimate to mean anything beyond the four rovers and three landing sites that generated it. The lens estimates a contribution; it does not interpret it. That is the work of the second anchor, and the seam between the two is exactly where the dissertation's conceptual model is stitched. Leamer's reminder that econometric inference is fragile to specification choices [\[72\]](#ref-leamer2010) sits at this seam as a discipline on both lenses: it counsels that the estimate be reported across the nested specifications rather than at a single preferred one, so that the reader sees how the autonomy coefficient behaves as the hardware and terrain blocks enter, rather than being handed a single number whose robustness is asserted rather than shown.

## 2.4 The Mokyr lens: autonomy as prescriptive knowledge resting on a propositional base

### 2.4.1 The framework and its primary sources

This section argues that Mokyr's economic history of technology supplies the interpretive frame the estimation lens cannot, by explaining why a mechanical improvement and a software improvement should be expected to differ in the durability and transferability of their returns, and that this difference is not incidental to the dissertation but is the substantive content of what a confirmed software advantage would mean.

The primary source is The Gifts of Athena [\[85\]](#ref-mokyr2002), whose central analytical move is the distinction between two kinds of useful knowledge. Propositional knowledge is knowledge of natural regularities, the understanding of why something works: the physics of wheel-soil interaction, the geometry of stereo vision, the mathematics of path planning. Prescriptive knowledge is technique, the instructions that do the work: the design of a particular wheel, the code of a particular navigation routine. Mokyr's thesis is that sustained technological progress occurs only when prescriptive technique rests on a widening propositional base, because techniques discovered by trial without underlying theory tend to exhaust their local improvements and stagnate, while techniques that rest on deep understanding are extensible, because the understanding generates new techniques, and self-correcting, because the understanding explains why a technique failed and how to repair it. Mokyr develops the broader frame in his treatment of the political economy of technological change [\[84\]](#ref-mokyr1998), where he analyzes why societies resist or absorb innovation, and in his synthesis across twenty-five centuries of technological change [\[86\]](#ref-mokyr2013), which traces the long-run pattern in which knowledge-intensive techniques outpace craft refinement.

Two further sources fix the epistemology and the diffusion mechanism. Villoro's analysis of personal versus propositional knowledge [\[113\]](#ref-villoro1998) sharpens the propositional category by distinguishing knowledge-that from knowledge-by-acquaintance, which matters here because the autonomy channel's advantage rests specifically on codified, transmissible knowledge-that, the kind that can be written into flight software and uploaded, rather than on tacit operator skill. The work of O'Rourke, Rahman, and Taylor on trade, knowledge, and the Industrial Revolution [\[92\]](#ref-orourke2007) and on the demographic consequences of skill-biased mechanization [\[93\]](#ref-orourke2008) supplies the formal endogenous-growth machinery in which a widening knowledge base raises the returns to further knowledge, and Braunerhjelm and colleagues [\[21\]](#ref-braunerhjelm2010) supply the knowledge-diffusion-and-entrepreneurship bridge that explains how propositional advances become deployed technique. The interpretation this dissertation places on this body of work is that it gives a principled, pre-data reason to expect the two rover channels to differ, which is precisely what converts an estimate into a meaning.

### 2.4.2 Why a wheel is a finished technique with bounded returns

Mechanical platform improvement is, in Mokyr's terms, prescriptive refinement of a technique whose propositional base is closed, so its marginal returns are real but bounded. The physics governing a rover wheel, terramechanics, the mechanics of wheel-soil interaction, soil shear strength, slip-sinkage, is mature and largely settled: the propositional knowledge underlying a wheel is not expanding in a way that would generate qualitatively new wheels. A larger wheel, a stronger actuator, a heavier chassis are refinements of an established technique whose returns are bounded by the same physics that bounded the previous design. This follows from Mokyr's general result that techniques resting on a closed propositional base exhibit diminishing returns to refinement [\[85\]](#ref-mokyr2002), a pattern visible in the long historical record of mature mechanical technologies in which incremental refinement yields steadily smaller gains, which Mokyr documents across centuries [\[86\]](#ref-mokyr2013).

The mechanism by which this bounds the hardware channel's returns is nameable. The driver is that the propositional base underlying the wheel is not widening; the mechanism is that without new understanding, improvement is confined to scaling and material substitution within known physics; the observable effect is that each mechanical increment costs more mass and power for less marginal productivity; the operational consequence is that the hardware lever's gains are paid in full at launch and frozen at the design that left Earth; the strategic implication is that a productivity strategy resting on the mechanical channel faces a rising cost curve and no retrofit option. This argument concerns the rate of return to further mechanical refinement, not the level: mechanical platform gains across the fleet were real and substantial, and the dissertation's null hypothesis H0 takes seriously the possibility that they dominate. Mokyr's framework does not deny the reality of the mechanical gains; it predicts their boundedness going forward. Confidence in this characterization is moderate to high as applied theory: terramechanics is a mature field, but the claim that no qualitatively new wheel technology will emerge is a conjecture about the future, downgraded accordingly.

### 2.4.3 Why autonomy rests on a widening propositional base and is extensible

Autonomy software, by contrast, is the kind of improvement that rests on a deepening propositional base in perception, planning, and terrain modeling, and is therefore extensible in a way a wheel is not, and retrofittable in a way a wheel cannot be. The trajectory the domain literature documents bears this out: the propositional knowledge underlying onboard autonomy, stereo perception, visual odometry, real-time path planning, hazard modeling, has expanded substantially across the rover generations and continues to expand on an active research frontier. Maimone, Cheng, and Matthies [\[79\]](#ref-maimone2007) document visual odometry as a capability that rested on advances in feature tracking and pose estimation, knowledge that was still developing. Biesiadecki, Leger, and Maimone [\[20\]](#ref-biesiadecki2007) analyze the directed-versus-autonomous driving tradeoff as one that shifted as the software's planning capability deepened. Verma et al. [\[112\]](#ref-verma2025) describe Enhanced Navigation as a capability that became possible because perception and planning advanced to the point that imagery could be processed and paths planned while the rover was in motion, a change in technique that rode directly on a widening propositional base. This is what Mokyr's result predicts, that techniques on a widening propositional base remain extensible and can sustain high marginal returns [\[85\]](#ref-mokyr2002), reinforced by the endogenous-growth formalization in which a deeper knowledge base raises the productivity of further knowledge work [\[92\]](#ref-orourke2007), [\[21\]](#ref-braunerhjelm2010).

The mechanism is the mirror of the hardware case. The driver is that the propositional base underlying autonomy is widening on an active research frontier; the mechanism is that new understanding in perception and planning generates qualitatively new techniques rather than mere scalings of old ones; the observable effect is that successive autonomy generations deliver step changes in autonomous-drive fraction and meters per sol rather than diminishing increments; the operational consequence, and this is the feature with no mechanical analogue, is that a new technique resting on transmissible propositional knowledge can be coded and uploaded to a rover already on the surface, because the knowledge is knowledge-that and not embodied in hardware [\[113\]](#ref-villoro1998); the strategic implication is that the software lever's cost is paid in development and verification rather than at launch, and its gains can reach deployed systems. Extensibility is a tendency, not a guarantee: the propositional base could mature and the software channel could itself reach diminishing returns, and the retrofit option is constrained by flight-software verification cost and by the fixed flight processor, which cannot be upgraded after launch even though the software it runs can. Confidence is moderate: the historical pattern and the domain evidence both point this way, but the claim is a structured conjecture about returns, not an established regularity, and the dissertation labels it as such.

### 2.4.4 The access-cost condition

Mokyr's framework attaches a condition to the persistence of knowledge-driven gains, that the access cost between research and deployment must stay low, and this condition maps onto a specific, observable feature of the Mars program: whether autonomy advances actually reach the surface through flight-software updates or are frozen at launch. Mokyr argues that propositional advances translate into prescriptive gains only when the cost of accessing and deploying the new knowledge is low; high access costs sever the link and leave understanding without productivity payoff [\[85\]](#ref-mokyr2002), [\[21\]](#ref-braunerhjelm2010). The underlying diffusion logic is that knowledge which cannot be cheaply transmitted from research to application does not raise productivity even when it exists. The documented Mars-program practice supports the low-access-cost case directly: flight-software updates deliver new mobility capability to rovers already on the surface, the clearest case being the in-flight evolution of navigation capability across a mission [\[112\]](#ref-verma2025), [\[50\]](#ref-gao2021).

The transfer is that the access-cost condition is not merely a theoretical caveat but a measurable moderator of the dissertation's central claim. If autonomy advances reach the surface through updates, the software channel's gains are realized on deployed rovers and the access cost is low; if they are frozen at launch because verification is prohibitive or the processor cannot run them, the channel's theoretical extensibility does not translate into realized productivity. One caveat is that the dissertation's panel measures realized productivity, so a low access cost is partly assumed by the very observation of within-mission capability growth, and the design must be careful not to treat the access-cost condition as both assumption and conclusion. Confidence is moderate: the Mars program demonstrably does upload mobility-relevant software, which supports a low access cost, but quantifying the access cost precisely is beyond the public archive and is flagged as an interpretive limit rather than a measured quantity.

### 2.4.5 What the lens does and does not provide

This section advances the interpretive half of the dissertation's argument: that a recovered software advantage would be material and durable rather than a local artifact. The Mokyr lens supports that by supplying the propositional-prescriptive distinction that predicts which channel's gains persist, the extensibility-and-retrofit argument that gives the software channel an operational advantage with no mechanical analogue, and the access-cost condition that names what must hold for the advantage to be realized. What the lens does not provide is any way to establish, from the data, that the software channel actually dominates; it predicts a pattern but cannot estimate it. That is the work of the first anchor. The two sections have now each delivered exactly half of what the dissertation needs, and the next section joins them.

## 2.5 The integrated conceptual model the empirical work will test

### 2.5.1 The model in one statement
This section argues that the two anchors compose into a single conceptual model with a testable core and an interpretive shell, and that the dissertation's design operationalizes that model. The model is this: the durable rise in Mars surface mobility productivity is driven principally by the autonomy-software generation and onboard hazard-detection capability rather than by mechanical platform growth (the substantive proposition); this is established by a terrain- and hardware-conditioned panel with a within-rover autonomous-fraction identification (the Angrist-Pischke core); and it is interpreted as an instance of extensible, propositional-base-resting technique outpacing bounded prescriptive refinement, with the access-cost condition determining whether the advantage is realized on deployed systems (the Mokyr shell). The core is what gets tested. The shell gives the test result its meaning and its forward implication.

The model's epistemic status must be stated cleanly. The substantive proposition is the dissertation's H1, reproduced here in the fixed operative language from the shared design: conditional on terrain class and on the same hardware covariates, autonomy-software generation and onboard hazard-detection capability explain the larger share of the variation in per-sol mobility productivity; the autonomy-generation coefficients are jointly significant and remain so after hardware covariates are included, and they account for more of the between-rover and within-rover variance than the hardware block does. The null H0 is its complement, also in fixed language: conditional on terrain class, cross-rover and within-rover variation in per-sol mobility productivity is explained by mechanical platform covariates, wheel diameter, actuator and mass class, available drive energy, and not by autonomy-software generation, with the autonomy-generation coefficients jointly indistinguishable from zero once hardware covariates are included. The model does not assume H1. It specifies the comparison and the falsification checks that would let the data choose between H1 and H0.

### 2.5.2 How the Angrist-Pischke lens maps onto the design

The mapping from the estimation lens to the design is one-to-one, and it is the reason the design takes the form it does. The named-counterfactual discipline maps onto the choice of the within-rover autonomous-drive-fraction contrast as the primary identification: this is the comparison that approximates the counterfactual of the same rover on the same terrain under a different amount of software involvement, with the mechanical platform held fixed at the level of the individual machine. The bad-controls discipline maps onto the rule that realized slip and realized wheel-soil interaction are mediators that stay on the outcome side, never entering the right-hand side, while a priori terrain class and commanded-drive parameters are eligible controls. The few-cluster discipline maps onto the wild-cluster bootstrap inference for the between-rover contrasts and onto the decision to make the within-rover contrast primary, since it sidesteps the small-cluster and heterogeneous-weighting problems the between-rover contrast faces [\[29\]](#ref-chaisemartin2020), [\[54\]](#ref-goodmanbacon2021). The credibility discipline maps onto the three pre-specified falsification checks. Each discipline becomes a concrete design choice, and the chapter's claim that the lens transfers without distortion is cashed out as this explicit correspondence rather than asserted.

The notation the design uses is fixed and is reproduced here so the conceptual model is anchored to the estimator the later chapters develop. For drive-sol *i* on rover *r* in terrain class *c*, productivity is modeled as a function of an autonomy-generation term, a hardware term, a terrain term, a rover fixed effect, and a terrain-class fixed effect:

\[
\text{Productivity}_{irc} = \beta_1 \, \text{AutonomyGen}_{r}
                 + \beta_2 \, \text{Hardware}_{r}
                 + \gamma \, \text{Terrain}_{ic}
                 + \alpha_{r} \quad \text{(rover fixed effect)}
                 + \delta_{c} \quad \text{(terrain-class fixed effect)}
                 + \epsilon_{irc} \qquad\qquad (1)
\]

The rover fixed effect \(\alpha_{r}\) absorbs fixed rover-specific characteristics, and the terrain-class fixed effect \(\delta_{c}\) absorbs systematic between-terrain-class differences. The tension the conceptual model confronts is that autonomy generation and hardware class are both rover-level attributes, so a rover fixed effect that fully absorbs rover identity also absorbs the categorical autonomy generation. The model resolves this exactly as the design does: the continuous autonomous-drive fraction varies within rover, because the same machine drives some segments autonomously and others under blind command, and this within-rover variation identifies the autonomy effect even with rover fixed effects in place, while the categorical generation contrast is identified from between-rover variation by relaxing the rover fixed effect to a terrain-class fixed effect plus the hardware block. The two estimates bound the effect. The within-rover estimate is conservative and clean because hardware is constant at the machine level; the between-rover estimate is broader in scope but leans on the hardware controls to separate the collinear channels.

### 2.5.3 How the Mokyr lens maps onto the interpretation

The mapping from the interpretation lens to the dissertation is equally specific. The propositional-prescriptive distinction maps onto the substantive reading of whichever coefficient dominates: a dominant autonomy block reads as extensible technique on a widening propositional base, a dominant hardware block reads as bounded prescriptive refinement of a closed technique. The extensibility-and-retrofit argument maps onto the dissertation's central policy claim, developed in Chapter 7, that if the autonomy channel dominates then productivity is bought more cheaply with software than with mass, the cost is paid in development rather than at launch, and the capability can be uploaded to a rover already on the surface. The access-cost condition maps onto the caveat that this policy advantage is realized only where flight-software updates actually reach the surface, which the Mars program's practice supports but which the panel cannot fully quantify. The model thus uses Mokyr not as decoration but as the source of the forward-looking implications that make the estimate decision-relevant rather than merely descriptive.

The named causal mechanism that the conceptual model proposes, and that the empirical work will test through the terrain-interaction prediction, runs from driver to strategic implication in a single chain. The driver is the advance in autonomy-software generation, G1 to G2 to G3, culminating in Enhanced Navigation that processes imagery and plans while driving. The mechanism is that a larger share of each drive is executed under onboard hazard detection and real-time path planning rather than blind commanded motion. The observable effect is a higher autonomous-drive fraction and more localized meters per sol, concentrated in hazard-dense terrain classes where onboard hazard detection most reduces the blind-driving penalty. The operational consequence is fewer sols consumed per meter, freeing sols for sampling, instrument placement, and caching in the daily planning cycle. The strategic implication is that surface mobility productivity is bought more cheaply with software than with mass, the cost is paid in development rather than at launch, and the capability can be uploaded to a rover already on the surface. This chain is not a correlation dressed as a cause: each link names a specific transmission, and the terrain-interaction prediction, that the effect should be largest where hazards are densest, is the link the data can test and the falsification check can refute.

### 2.5.4 The model's testable implications and their illustrative expected directions

The conceptual model yields three testable implications, each of which the empirical chapters operationalize and each of which can fail. The implications are stated here with their expected directions, and every expected value is labeled illustrative, not estimated, because this is a design-stage dissertation and no panel has yet been estimated.

The first implication concerns the nested decomposition. The model predicts that adding the autonomy block to a specification already containing the hardware and terrain blocks absorbs a larger incremental share of variance than the hardware block absorbs, and that the hardware coefficients attenuate when the autonomy block enters, indicating that part of what looked mechanical was the correlated autonomy upgrade. Illustratively, and not as an estimate, the hardware block alone is expected to absorb a moderate share of between-rover variance consistent with real mechanical gains, and the autonomy block is expected to absorb a larger incremental share. If instead the autonomy block adds little incremental share, the first falsification check fires and the model's core proposition is in doubt.

The second implication concerns the within-rover contrast, which is the model's primary test. The model predicts that the autonomous-drive-fraction coefficient is positive and survives the inclusion of rover and terrain-class fixed effects, because within a single machine the only thing differing between autonomous-heavy and blind-commanded segments is the software's contribution. Illustratively, and not as an estimate, this coefficient is expected to be positive and to remain so after fixed effects. A within-rover coefficient indistinguishable from zero after terrain conditioning fires the strongest falsification check and would move the verdict toward H0.

The third implication concerns the terrain interaction, which is the model's mechanism test. The model predicts that the autonomy effect is largest in rougher, more hazard-dense terrain classes and smallest on benign flats where blind commanded driving is already near-optimal, because the named mechanism is that onboard hazard detection reduces the blind-driving penalty most where hazards are dense. Illustratively, and not as an estimate, the autonomy-by-terrain interaction is expected to be concentrated in the rough classes. A flat interaction pattern would leave the mechanism unsupported even if a raw autonomy correlation survived, which is why this check is specified as distinct from the first two: it tests not whether autonomy matters but whether it matters for the reason the model claims.

### 2.5.5 Confidence, residual risk, and what would move the verdict

The confidence the integrated model can carry is set out here, claim by claim and calibrated to the design-stage evidence. The convergent mission and survey literature establishes at high confidence that productivity rose steeply while the two channels were described together but never separated [\[39\]](#ref-crisp2003), [\[55\]](#ref-grotzinger2012), [\[112\]](#ref-verma2025), [\[50\]](#ref-gao2021), and the mission-campaign record establishes, also at high confidence, that sols are the binding scarcity for Mars Sample Return caching and instrument placement [\[45\]](#ref-farley2023), [\[52\]](#ref-golombek2014). The claim that the design reaches a causal mechanism rather than a bare correlation carries moderate-to-high confidence on the strength of the mapping in 2.5.2, tempered by the residual risk that the within-rover autonomous fraction is only approximately as-good-as-random after terrain conditioning. That the within-rover design improves on its rivals, the combined-narrative description and the naive between-rover regression, likewise carries moderate-to-high confidence, supported by the heterogeneous-effects literature that shows why the naive estimator misbehaves [\[29\]](#ref-chaisemartin2020), [\[54\]](#ref-goodmanbacon2021), [\[23\]](#ref-callawaysantanna2021). The remaining identification risk is held at moderate confidence, because few-cluster inference and explicit threat treatment bound it without dissolving it, and three rovers will never be many.

The evidence that would raise confidence is specific and is named so the design is honest about its own limits. A clean, published per-drive autonomous-fraction series across all three generations would raise confidence in the within-rover test from moderate-to-high toward high, and its current incompleteness in the public archive is the single highest-value gap to close before execution; the design must reconstruct the fraction and carry the measurement-error caveat until that series exists. A terrain-class crosswalk that ties each drive to a class on independent orbital grounds, rather than on the analyst's construction, would raise confidence in the mechanism test. A demonstration that the autonomous fraction's residual variation is uncorrelated with productivity-relevant terrain features after conditioning would raise confidence in the identification itself. Evidence that lowers confidence is equally nameable: a strong correlation between the reconstructed autonomous fraction and unobserved terrain difficulty would weaken the within-rover test; a finding that the flight processor, not the software, gated the generation step would partly merge the two channels the model seeks to separate. The model is therefore not asserted. It is positioned with its confidence calibrated to the design-stage evidence grade, and the chapters that follow build the data, the design, and the analysis that would let the data deliver the verdict the model anticipates but does not presume.

### 2.5.6 Why two anchors and not one, restated as the chapter's close

The chapter opened by claiming that the two anchors do non-substitutable work, and it closes by showing that the integrated model is the proof of that claim. Strip out Angrist and Pischke and the model loses its core: there is no within-rover comparison, no bad-controls rule, no few-cluster honesty, and the substantive proposition reverts to the combined-narrative description the dissertation set out to improve. Strip out Mokyr and the model loses its shell: the estimate, even if recovered cleanly, would carry no implication for whether the advantage recurs in the next rover, transfers to a lunar platform, or can be uploaded after landing, and the policy claim that gives the dissertation its purpose would have no theoretical grounding. The estimation lens establishes a contribution it cannot interpret; the interpretation lens predicts a pattern it cannot establish. The conceptual model is the join, and the dissertation's value is precisely that it refuses to settle for either half. The next chapter turns to the domain literature, which is rich in the description both anchors discipline and thin in the identification the first anchor supplies, and which therefore constitutes both the evidentiary base for the model's premises and the gap the model is built to fill.



# Chapter 3: Literature Review

## 3.0 Chapter thesis and orientation

The domain literature on Mars surface mobility is rich in description and thin in identification. It establishes, credibly and repeatedly, that each successive flight rover was at once a better machine and a smarter one, and that per-sol traverse productivity rose steeply across the fleet from Sojourner through the Mars Exploration Rovers, Curiosity, and Perseverance. What the literature never does is construct the comparison that would isolate either causal channel from the other. The mechanical platform and the autonomy-software generation were upgraded together on every mission, so the two improvements arrive bundled and are reported bundled, and no published study holds terrain and hardware fixed in order to ask how much of the productivity gain is attributable to the autonomy-software generation alone. That missing comparison is the gap this dissertation fills, and the purpose of this chapter is to demonstrate, source by source, that the gap is real, that it is material, and that the existing evidence base supplies every ingredient needed to close it except the identification design itself.

The argument of the chapter is therefore not that the literature is wrong but that it answers a different question than the one a mission designer must answer. The space-robotics literature answers "what did each rover do and how did it work?" with great fidelity. It does not answer "if a fixed mass-and-power budget could be spent either on heavier mechanical capability or on more onboard computation, which buys more meters per sol?" The first question is descriptive and the second is causal, and the distance between them is exactly the distance between a narrative that credits the whole machine and a design that partitions the credit. The methodological anchors of this dissertation name that distance precisely: design-based econometrics insists that a causal claim must specify the counterfactual and the comparison that identifies it (Angrist and Pischke [\[7\]](#ref-angrist2009)), and the economic history of technology insists that the durability of a technical gain depends on whether it rests on a finished technique or on a still-widening base of understanding (Mokyr [\[85\]](#ref-mokyr2002)). This chapter reads the domain literature through both lenses, organizes it by the autonomy generation it documents, and closes with an explicit gap statement and the propositions that follow.
The problem this chapter addresses can be framed in four moves. The current state of knowledge is a large descriptive literature in which the autonomy and hardware channels are always discussed together and never separated. The desired state is a body of evidence that, assembled into a panel, supports a clean partition of the productivity gain between the two channels. The gap is the absence of any identification strategy in the existing work: collinearity of the two channels is acknowledged implicitly by everyone and confronted explicitly by no one. Leaving the gap open carries a consequence. Mission planners continue to make a high-stakes, asymmetric allocation decision, mass versus software, on the strength of a combined narrative rather than a partitioned estimate. The asymmetry matters because a software shortfall can sometimes be patched after landing while a hardware shortfall cannot.

A note on confidence and calibration is owed at the outset. This is a design-stage dissertation, and the literature review must be honest about what the cited sources can and cannot support. Where a source reports a measured engineering result on a flight rover, the evidence grade is high and the claim resting on it is stated with high confidence. Where a source is a simulation, a single-wheel laboratory test, or a terrestrial-analog field trial, the evidence grade is lower and the inference to flight behavior is qualified accordingly. Where the dissertation's own propositions go beyond what any single source establishes, they are labeled as the gap-closing contribution and not as findings already in the literature. No empirical result is claimed here that the assembled panel has not yet produced, and every forward-looking expectation is marked as such.

## 3.1 The Mars Exploration Rover autonomy baseline

The Mars Exploration Rover mission established the baseline against which all subsequent surface-mobility autonomy is measured, and the chapter therefore begins there. This section advances the case that the MER generation introduced three capabilities that are individually documented and jointly constitute the first coherent autonomy generation in the panel: onboard stereo hazard assessment, visual odometry for slip-aware position estimation, and global path planning layered on top of the local planner. The primary engineering reports describe capabilities that did not exist on the prior Sojourner microrover and that were inherited, recognizably, by the two later flight rovers, and the mission overview literature situates these capabilities in flight operations.

The mission and its design are documented in the foundational overview (Crisp et al. [\[39\]](#ref-crisp2003)), which describes the twin rovers' science payload, the stereo navigation and hazard camera suite, and the once-per-sol command cadence that makes onboard autonomy valuable in the first place. That framing matters for this dissertation because it makes explicit the operational scarcity that productivity is measured against: each rover was commanded once per Martian day on the basis of the previous sol's data, so any meter the rover could plan and execute without a ground decision was a meter that did not consume a planning cycle. The companion overview of the Opportunity rover's traverse from Eagle Crater to the Purgatory Ripple (Squyres et al. [\[106\]](#ref-squyres2006)) records the actual surface distances achieved, more than five kilometers of outcrop-scale investigation, and in doing so it documents the kind of cumulative traverse record that the PDS archives will later resolve into the per-sol productivity series this dissertation requires.

The autonomy substrate of the MER generation rests on stereo hazard assessment performed by the GESTALT local planner (grid-based estimation of surface traversability applied to local terrain), the system that lets the rover build a local terrain map from its stereo cameras and reject unsafe steps without a ground-in-the-loop decision. The most consequential single capability was visual odometry. The definitive engineering account reports two years of visual odometry operations across Spirit and Opportunity (Maimone, Cheng, and Matthies [\[79\]](#ref-maimone2007)), and its findings are precise enough to serve as load-bearing evidence here. The system computed a six-degree-of-freedom pose update by tracking autonomously selected terrain features between stereo pairs, converged successfully on the large majority of attempts (the report gives 97 percent on Spirit and 95 percent on Opportunity), detected slip ratios as high as 125 percent, and measured changes as small as two millimeters while driving on slopes as steep as 31 degrees. The method was image-based feature tracking with a maximum-likelihood estimator over computed 3D offsets; the limitation, stated in the source, is that visual odometry was computationally expensive on the flight processor and was therefore used selectively rather than continuously. The interpretation for this dissertation is direct. Visual odometry is the mechanism by which the MER generation converted otherwise-dangerous sloped and sandy terrain into traversable terrain. It is a software capability that raised productivity on exactly the terrain classes where the hardware platform alone would have stalled or driven blind into a slip failure.

The earlier conference treatment of the same system (Cheng, Maimone, and Matthies [\[34\]](#ref-cheng2006)) supplies the design-goal context that the journal article assumes: onboard estimates from the inertial measurement unit and wheel encoders alone met the design goal of at most 10 percent error in benign terrain, but on slopes tilted as high as 31 degrees that dead-reckoned estimate degraded, and visual odometry was the additional capability invoked to maintain position knowledge. The retrospective magazine treatment (Yang, Maimone, and Matthies [\[121\]](#ref-yang2006)) adds the operational lesson that matters most for the productivity question: visual odometry achieved difficult drive approaches in fewer sols and ensured accurate science imaging, but it required active pointing by human drivers in feature-poor terrain. That dependence is itself evidence for the dissertation's treatment of the autonomous-drive fraction as a within-rover variable, because it shows that even within a single machine the share of a drive executed under the software's contribution varied with terrain and with ground decisions. The attitude- and position-estimation account (Ali et al. [\[5\]](#ref-ali2006)) completes the picture by documenting how the rovers acquired attitude from accelerometers and sun images and propagated it with gyros and wheel odometry, with visual odometry invoked to fine-tune position updates specifically in high-slip environments.

The trade between blind commanded driving and autonomous hazard-avoidance driving is the single most important piece of prior work for this dissertation's identification strategy, and it is documented directly (Biesiadecki, Leger, and Maimone [\[20\]](#ref-biesiadecki2007)). The report describes how MER mobility operations were encoded as event-driven sequences and how driving strategies "alternately use more or less onboard autonomy, to maximize drive speed and distance at the cost of increased complexity in the sequences of commands built by human Rover Planners each day." The method is operational analysis of flown drives; the finding is that solar energy generally permitted at most about four hours of driving per sol and that drive time was further restricted by other planned activities, so that the choice between directed and autonomous driving was a real, recurring, terrain-contingent decision. The limitation is that the report is descriptive and does not estimate the productivity differential between the two modes in a way that controls for terrain. That precise gap, a documented mode choice that varies within rover, never converted into a terrain-conditioned estimate of its productivity effect, is the opening this dissertation's within-rover autonomous-fraction contrast is designed to exploit. The bad-controls discipline (Angrist and Pischke [\[7\]](#ref-angrist2009)) is what keeps the estimate honest, because the report makes clear that realized slip and realized drive time are produced by the drive rather than fixed before it.

The path-following work under high slip (Helmick et al. [\[59\]](#ref-helmick2004)) supplies the failure mode that most directly couples terrain to productivity. It demonstrates path following using visual odometry for a Mars rover in high-slip environments, the conditions under which wheel odometry alone is dangerously wrong and the rover would otherwise have to drive conservatively or not at all. The method is a control architecture that closes the loop on visually estimated position; the relevance is that it identifies the specific terrain-software interaction (high slip plus onboard slip estimation) that the dissertation predicts will show the largest autonomy effect.

The extension of the MER stack from purely local to global path planning is documented in two paired sources (Carsten, Rankin, Ferguson, and Stentz [\[26\]](#ref-carsten2007); Carsten, Rankin, Ferguson, and Stentz [\[27\]](#ref-carsten2009)). The conference paper and the journal article describe the integration of a Field D* global planner into MER flight software, enabling simultaneous local and global planning during AutoNav. The motivating finding, stated plainly, is that the GESTALT local planner "works well at guiding the rovers around narrow and isolated hazards; however, it is susceptible to failure when clusters of closely spaced, nontraversable rocks form extended obstacles." The method was software integration followed by surface testing on the flight vehicles; the result was a planning capability that extended the rover's effective planning horizon from the immediate local patch to a larger map and reduced the frequency with which a locally greedy choice led into a cul-de-sac. The limitation is that the global planner's benefit is concentrated in cluttered, extended-obstacle terrain and is near-zero on open ground, which again is a terrain-conditioned effect rather than a uniform one. The interpretation for this dissertation is that the MER autonomy generation is not a single capability but a stack (hazard assessment, visual odometry, global planning) whose productivity payoff is terrain-dependent in a way the design must model through the terrain-class fixed effect and the autonomy-by-terrain interaction.

Two strands of pre-flight prototype work establish that these capabilities were the product of a deepening research base rather than ad hoc flight expedients, which matters for the Mokyr reading developed in the synthesis. The Rocky 7 sciencecraft prototype (Hayati et al. [\[58\]](#ref-hayati2002)) demonstrated long traverses, autonomous navigation, and science-instrument control in Mojave Desert field tests, and the sun-sensor heading-determination results (Volpe [\[114\]](#ref-volpe2003)) reported cross-track error of only 6 percent of distance traveled over a one-kilometer field traverse, an improvement over prior planetary-rover systems. The European-lineage autonomous-navigation work (Maurette [\[81\]](#ref-maurette2003)) documents a parallel development track. These prototype results are graded as analog evidence (terrestrial field trials, not flight), and the inference from them to flight productivity is qualified accordingly, but they are the propositional-base evidence Mokyr's framework predicts should underlie an extensible technique.

The following synthesis table organizes the MER-generation evidence by capability, method, finding, and the limitation that bears on this dissertation's gap.

| Capability (MER / G1) | Representative source(s) | Method | Key finding | Limitation for the productivity question |
|---|---|---|---|---|
| Stereo hazard assessment (GESTALT) | Crisp et al. [\[39\]](#ref-crisp2003); Carsten et al. [\[27\]](#ref-carsten2009) | Onboard local traversability grid from stereo | Rejects unsafe steps without ground decision; fails on extended obstacle clusters | Benefit is terrain-contingent; no terrain-conditioned productivity estimate |
| Visual odometry | Maimone, Cheng, Matthies [\[79\]](#ref-maimone2007); Cheng et al. [\[34\]](#ref-cheng2006); Ali et al. [\[5\]](#ref-ali2006) | Feature tracking across stereo pairs, ML pose update | 95-97% convergence; slip detected to 125%; safe driving on 31-degree slopes | Computationally expensive, used selectively; productivity gain not isolated from hardware |
| Directed vs autonomous mode choice | Biesiadecki, Leger, Maimone [\[20\]](#ref-biesiadecki2007) | Operational analysis of flown drives | Mode chosen daily to trade speed/distance against command complexity | Descriptive; no terrain-controlled estimate of the productivity differential |
| Global path planning (Field D*) | Carsten et al. [\[26\]](#ref-carsten2007), [\[27\]](#ref-carsten2009) | Field D* integrated into flight software, surface-tested | Extends planning horizon; reduces cul-de-sac failures in cluttered terrain | Benefit concentrated in extended-obstacle terrain only |
| Path following in high slip | Helmick et al. [\[59\]](#ref-helmick2004) | Visual-odometry-closed control loop | Enables driving where wheel odometry alone is unsafe | Identifies the terrain-software interaction but does not quantify its fleet-level effect |

The convergence across these sources warrants a high-confidence claim that the MER generation constitutes a genuine, internally documented autonomy generation (G1 in the dissertation's notation) whose productivity contribution is real, terrain-dependent, and never separated in the literature from the contemporaneous mechanical platform. That last clause is the gap in miniature.

## 3.2 The Mars Science Laboratory generation: inheritance, extension, and the terrain ceiling

This section argues that the Mars Science Laboratory generation (Curiosity, G2 in the dissertation's notation) inherited the MER autonomy stack and extended it onto a much larger and heavier platform, and that the operational lesson of the mission was that terrain interaction, not onboard intelligence, set the productivity ceiling on the worst ground. The mission overviews and the wheel-damage and megaripple-crossing analyses document a rover whose autonomy was recognizably descended from MER while its mechanical envelope and its terrain hazards changed substantially, and the terramechanics literature explains why the larger platform's wheels became the binding constraint.

The mission is established in its foundational overview (Grotzinger et al. [\[55\]](#ref-grotzinger2012)) and its first-500-sols operational account (Vasavada et al. [\[110\]](#ref-vasavada2014)). The latter is particularly valuable for this dissertation because it documents the operational reality that productivity is measured against: Curiosity traversed eastward through Gale crater investigating outcrops and sampling aeolian and lacustrine deposits, and "the unprecedented complexity of the rover and payload systems posed challenges to science operations, as did a number of anomalies," with operational processes revised to add advance-planning opportunities. This is direct evidence that the productivity series for G2 is shaped not only by autonomy and hardware but by an evolving ground-operations practice, which is precisely the operational-learning rival the dissertation must separate from the onboard-autonomy channel using the autonomous-drive-fraction measure.

The autonomy continuity is documented at the algorithmic level. The visual-odometry feature-tracking work developed for Mars Science Laboratory (Johnson et al. [\[67\]](#ref-johnson2008)) reports an algorithm at least four times more computationally efficient than the MER version while tracking significantly more features, validated on thousands of terrestrial and Martian stereo pairs. The method is an integrated motion-estimation and stereo-feature-tracking loop; the finding is a substantial efficiency and robustness improvement over the MER baseline. The interpretation is that G2 is not a relabeling of G1 but a genuine extension resting on a deeper algorithmic base, the kind of distinction the dissertation's TechPort robustness check is designed to verify externally and which the Mokyr lens reads as prescriptive improvement resting on a widening propositional base.

The defining empirical contribution of the MSL generation for this dissertation is the terrain literature, because it establishes that on the worst ground the binding constraint was mechanical and terramechanical rather than computational. The megaripple-crossing analysis up to sol 710 (Arvidson et al. [\[18\]](#ref-arvidson2017)) is the central source. Its method combined imaging and engineering data from flown traverses with laboratory single-wheel soil experiments and full-scale rover testing. Its findings are quantitatively specific and load-bearing: traverses across megaripple deposits produced sinkage up to approximately 30 percent of the 0.50-meter wheel diameter, high compaction resistances, and rover-based slip up to 77 percent. The companion wheel-damage analysis (Arvidson et al. [\[19\]](#ref-arvidson2017b)) documents that early drives across sharp sandstone outcrops produced an unacceptably high rate of punctures and cracks in the thin aluminum wheel skins, that the damage was tied to the drive-control mode and the rocker-bogie kinematics (wheels leading a suspension pivot were forced onto sharp immobile surfaces by the trailing wheels), and that a geomorphic map was generated to plan traverses minimizing further damage. The method is forensic engineering analysis tied to geologic mapping; the finding is that terrain hazard drove a change in routing and driving practice independent of the rover's intelligence. The interpretation for this dissertation is unambiguous and high-confidence: on the roughest MSL terrain, productivity was throttled by wheel-soil interaction and wheel survivability, which are mechanical and terramechanical phenomena, so the terrain-class fixed effect and the mediator treatment of realized slip are not refinements but necessities. Realized slip up to 77 percent is exactly the kind of post-treatment, drive-produced quantity that the bad-controls discipline forbids on the right-hand side (Angrist and Pischke [\[7\]](#ref-angrist2009)), because conditioning on it would absorb part of whatever autonomy effect exists.

The terrain-characterization literature supplies the covariate basis. The terrain physical properties derived from the first 360 sols (Golombek et al. [\[52\]](#ref-golombek2014)) integrate orbital data with rover-based measurements of sinkage, slip, and traverse behavior to classify the hummocky-plains and bedded-fractured units the rover crossed. This is the source from which a terrain-class crosswalk is constructed; the dissertation flags in Chapter 4 that the per-drive class assignment is built by the candidate from these characterizations rather than cited ready-made, which is a construction step, not a citable result. The supporting geological and compositional literature on the Curiosity traverse (the Bagnold Dunes spectroscopy of Lapôtre et al. [\[71\]](#ref-laptre2017); the extended-mission geologic mapping of Stack et al. [\[14\]](#ref-anon2016b); the diagenesis and mobility studies of L'Haridon [\[64\]](#ref-j2020), Berger [\[65\]](#ref-ja2022), Haber et al. [\[57\]](#ref-haber2025); the methane-transport modeling of Pla-Garcia et al. [\[98\]](#ref-plagarcia2024)) is graded as context rather than as direct evidence on mobility productivity. These sources establish where the rover went and what it found, which fixes the terrain and the cumulative-distance record, but they do not speak to the autonomy-versus-hardware partition and are cited as traverse-context provenance rather than as mobility evidence.

The traverse-simulation literature deserves separate attention because it represents the field's own attempt to predict per-sol mobility, and its method illuminates both the value and the limits of a model-based approach to the question this dissertation answers empirically. The simulation of Mars rover traverses (Genova et al. [\[51\]](#ref-genova2013)) constructs forward models of rover progress over characterized terrain; the soft-soil parameter-uncertainty work (Gallina, Krenn, and Schäfer [\[49\]](#ref-gallina2016)) quantifies how uncertainty in the underlying soil parameters propagates into mobility-simulation predictions. The method in each is a physics-based or semi-empirical forward simulation; the finding is that per-sol progress can be predicted to a useful approximation given terrain and platform parameters, but only with explicit uncertainty bounds. The interpretation for this dissertation is twofold and carefully calibrated. First, these simulations confirm that terrain and platform together determine a large share of productivity, the mechanical channel the design must condition on rather than dismiss. Second, a forward simulation cannot by itself partition the observed productivity gain between the autonomy and hardware channels, because the simulation takes the platform and the driving policy as inputs rather than estimating their separate contributions from observed variation. This is the structural difference between a model that predicts and a design that identifies (Angrist and Pischke [\[7\]](#ref-angrist2009)): the simulation literature is descriptive-predictive, and the dissertation's contribution is identificatory. The two are complements, not substitutes, and the simulation work is cited here to mark the boundary precisely.

The operational-learning rival deserves explicit treatment in this section because the MSL literature is where it is most clearly visible and most readily confused with the autonomy channel. The first-500-sols account (Vasavada et al. [\[110\]](#ref-vasavada2014)) documents that ground-operations processes were revised over the mission to add advance-planning opportunities, and the wheel-damage analysis (Arvidson et al. [\[19\]](#ref-arvidson2017b)) documents that a geomorphic map was generated specifically to let human planners route around hazards. Both are improvements in how the ground team plans drives, and both raise productivity over the life of the mission. The rival hypothesis they generate is that the upward productivity trend across the fleet reflects twenty years of human operational learning rather than onboard autonomy. The literature does not separate these, and indeed the wheel-damage source treats careful human path planning and the rover's own capabilities as jointly responsible for extending wheel life. The dissertation's response, anticipated here so the gap statement is complete, is that the autonomous-drive-fraction measure is the only construct that distinguishes the two, because ground learning improves blind-commanded drives as well as autonomous ones, whereas the autonomous-drive fraction is specifically the share executed onboard. This is why the within-rover autonomous-fraction contrast, rather than the cross-rover trend, carries the identifying weight, and it is a point the MSL literature makes unavoidable rather than optional.

The following synthesis table organizes the MSL-generation evidence.

| Theme (MSL / G2) | Representative source(s) | Method | Key finding | Bearing on the gap |
|---|---|---|---|---|
| Mission and operations | Grotzinger et al. [\[55\]](#ref-grotzinger2012); Vasavada et al. [\[110\]](#ref-vasavada2014) | Mission overview, first-500-sols ops | Larger platform; evolving ground-ops practice; anomalies | Operational learning is a rival channel to separate |
| Autonomy continuity/extension | Johnson et al. [\[67\]](#ref-johnson2008) | Efficient stereo feature tracking | 4x faster, more features than MER VO | G2 is a real extension, not a relabel; supports the generation construct |
| Terrain as productivity ceiling | Arvidson et al. [\[18\]](#ref-arvidson2017) | Flown traverses + single-wheel tests | Sinkage to 30% wheel diameter; slip to 77% on megaripples | Mechanical/terramechanical limit; mandates terrain FE and slip-as-mediator |
| Wheel survivability | Arvidson et al. [\[19\]](#ref-arvidson2017b) | Forensic engineering + geomorphic mapping | Sharp-rock punctures; routing changed to protect wheels | Hardware constraint shaping productivity independent of autonomy |
| Terrain covariate basis | Golombek et al. [\[52\]](#ref-golombek2014) | Orbital + rover sinkage/slip integration | Terrain physical-property classification | Source for the constructed terrain-class crosswalk |

The convergence here warrants a high-confidence claim that the MSL generation simultaneously advanced the autonomy stack and exposed a mechanical-terrain productivity ceiling, and that the literature treats the two as inseparable features of the same mission. The dissertation reads this not as a refutation of the autonomy channel but as the strongest available statement of the collinearity problem: on this rover, more than any other, the autonomy upgrade and the hardware-terrain reality are entangled, which is exactly why the within-rover autonomous-fraction contrast (hardware constant by construction) is the identification of choice rather than a between-rover regression.
## 3.3 The Mars 2020 generation: moving computation into the drive

This section argues that the Mars 2020 generation (Perseverance, G3) introduced the fleet's most consequential autonomy step by moving image processing and path planning into the drive itself, so that the rover no longer had to stop, think, and then move. That change is the most plausible single cause of the step increase in meters per sol on hazard-dense ground. The Enhanced Navigation engineering literature and the measured autonomous-distance records show that the timing of computation, not merely its quality, changed; the perception-substrate literature documents what made continuous in-drive planning feasible on a constrained flight processor.

The definitive account is the Enhanced Autonomous Navigation paper (Verma et al. [\[112\]](#ref-verma2025)). Its method is the design and flight characterization of ENav, the autonomous driving algorithm of Perseverance, on a single-core CPU with severely limited computing resources. Its central engineering contributions are a two-stage path-selection approach that balances path optimality against computational efficiency, a collision-checking algorithm that conservatively approximates expensive kinematic settling, and robustness against slip via expanded wheel bounding boxes. For this dissertation, ENav matters because it is the system that lets the rover plan while in motion, raising the autonomous drive fraction and the achievable distance per sol. The limitation, stated in the source, is that the achievement was constrained by the flight processor. The productivity gain therefore came from algorithmic timing and efficiency rather than from raw onboard computing power, a distinction the Mokyr lens reads as a knowledge-intensive rather than resource-intensive gain.

The measured outcomes are reported in the operational retrospective (Verma et al. [\[111\]](#ref-verma2023)), and they are the most directly relevant quantitative evidence in the entire corpus for the dissertation's dependent variable. AutoNav evaluated 88 percent of the 17.7 kilometers traveled during Perseverance's first Mars year. The prior maximum total autonomously evaluated distance had been 2.4 kilometers, by Opportunity over its 14-year lifetime. ENav set the records for greatest distance driven without human review (699.9 meters) and greatest single-day drive distance (347.7 meters), and the companion onboard scheduling and autonomous-science systems (the OnBoard Planner projected to reduce energy use by up to 20 percent; the AEGIS target-selection system) extended autonomy beyond driving into science. The method is operational reporting of flown distances; the finding is a step change in autonomous distance fraction between generations, from a 2.4-kilometer lifetime ceiling to 88 percent of a single Mars year. The interpretation, stated with the appropriate qualifier, is that this is the pattern H1 predicts: a sharp rise in the autonomous-drive fraction coinciding with the G3 autonomy generation. The source, however, reports it as an aggregate operational record and does not condition it on terrain or separate it from the contemporaneous hardware upgrade. The raw correlation is strongly suggestive; the causal partition is not in the source, and the dissertation says so. This is a correlation-not-causation boundary, and confidence in the causal reading is held at moderate pending the terrain- and hardware-conditioned estimation.

The perception and software substrate that made in-drive planning feasible is documented across several sources. The Mars 2020 engineering camera and microphone system (Maki et al. [\[80\]](#ref-maki2020)) describes a next-generation imaging system with 16 cameras, including 9 for surface operations (2 Navcams, 6 Hazcams, 1 Cachecam), with the Navcams acquiring color stereo at 0.33 milliradian per pixel over a 96-by-73-degree field of view. The method is instrument design and characterization; the relevance is that this richer, faster imaging substrate is the sensory input ENav consumes while driving. The machine-learning path-planning work for Enhanced AutoNav (Abcouwer et al. [\[3\]](#ref-abcouwer2021)) is the most algorithmically specific source. It documents that ENav sorts candidate paths and uses the Approximate Clearance Evaluation (ACE) algorithm to check safety, that ACE is computationally expensive, and that two heuristics (a Sobel-gradient cost and a machine-learned untraversability predictor) reduce the number of expensive ACE evaluations needed before a feasible path is found, with physics-simulation Monte Carlo trials quantifying the improvement. The method is heuristic design plus simulation evaluation; the finding is a concrete computational-efficiency mechanism for in-drive planning. The G3 productivity gain thus has a named, documented mechanism, cheaper safe-path evaluation enabling continuous planning, which is the kind of driver-to-mechanism-to-effect chain the dissertation requires for any causal claim rather than a bare correlation.

The pre-flight simulation infrastructure (Toupet et al. [\[108\]](#ref-toupet2020)) describes a ROS-based simulator for testing Enhanced Autonomous Navigation, evidence that the capability was validated in a deepening test environment before flight. The independent pose-estimation analyses of Perseverance's path (Di et al. [\[41\]](#ref-di2022); Andolfo, Petricca, and Genova [\[6\]](#ref-andolfo2022); Andolfo [\[101\]](#ref-s2021); Andolfo [\[102\]](#ref-s2022)) reconstruct the rover's localization on specific sols using stereo-vision visual odometry and find their reconstructed paths consistent with the telemetered onboard trajectory. The method is independent VO reconstruction validated against telemetry; the relevance is twofold. These analyses establish the accuracy of the localized-path-length measure that is the dissertation's dependent variable, and they demonstrate that the per-sol traverse products needed for the panel are recoverable and verifiable for the G3 rover.

The following synthesis table organizes the Mars 2020 evidence.

| Theme (M2020 / G3) | Representative source(s) | Method | Key finding | Bearing on the gap |
|---|---|---|---|---|
| Enhanced Navigation algorithm | Verma et al. [\[112\]](#ref-verma2025) | ENav design + flight characterization | Two-stage planning; in-drive computation on a single-core CPU | Names the mechanism: planning while moving raises autonomous fraction |
| Measured autonomous distance | Verma et al. [\[111\]](#ref-verma2023) | Operational distance reporting | 88% of 17.7 km autonomously evaluated; records of 699.9 m and 347.7 m | Strong raw correlation with G3; not terrain- or hardware-conditioned |
| Imaging substrate | Maki et al. [\[80\]](#ref-maki2020) | Camera-system design/characterization | 16 cameras; high-rate color stereo Navcams | Sensory input enabling continuous in-drive perception |
| Path-evaluation efficiency | Abcouwer et al. [\[3\]](#ref-abcouwer2021) | Heuristics + Monte Carlo simulation | Sobel + ML heuristics cut expensive ACE checks | Computational mechanism behind in-drive planning |
| Localization verification | Di et al. [\[41\]](#ref-di2022); Andolfo et al. [\[101\]](#ref-s2021), [\[102\]](#ref-s2022) | Independent stereo VO vs telemetry | Reconstructed paths consistent with onboard trajectory | Validates the localized-path-length dependent variable for G3 |

The convergence warrants a high-confidence claim that the G3 generation moved computation into the drive and a moderate-confidence claim that this change caused the observed step increase in autonomous distance. The latter is held at moderate precisely because the supporting distance figures are unconditioned aggregates. The dissertation's contribution is to convert the suggestive aggregate into a terrain- and hardware-conditioned estimate, and the literature surveyed here supplies both the mechanism (in-drive planning) and the measurement substrate (verified localized path lengths) needed to do so.

## 3.4 Terramechanics and wheel-soil interaction: the source of the hardware and terrain covariates

This section argues that the terramechanics literature supplies the physical basis for both the hardware covariate block and the terrain covariates, and that it independently establishes the dissertation's most important methodological commitment: realized slip and realized wheel-soil interaction are products of the drive and must be treated as mediators, never as right-hand-side controls. The wheel-soil interaction modeling literature models slip and sinkage as outcomes of commanded motion over terrain, which is exactly the structure the bad-controls discipline requires, and the breadth and convergence of the modeling tradition across analytical, experimental, discrete-element, and machine-learning methods give that structure independent weight.

The analytical foundation is the terramechanics tradition initiated for planetary rovers by the traction-control and steering-maneuver work (Ishigami et al. [\[61\]](#ref-ishigami2006); Ishigami et al. [\[63\]](#ref-ishigami2007); Ishigami, Nagatani, and Yoshida [\[62\]](#ref-ishigami2006b)). The traction-control analysis models wheel-soil interaction to evaluate traction and disturbance forces under varying slip; the steering-maneuver model decomposes traction into longitudinal and lateral components and identifies the side force as the dominant influence on steering on loose soil, validated against single-wheel test beds and a four-wheel rover on lunar-regolith simulant; the slip-compensation path-following work derives steering and driving maneuvers that simultaneously follow a path and compensate for slip. The method across these is semi-empirical modeling validated on test beds; the finding is that slip is a function of commanded motion, soil parameters, and slope, which is to say slip is downstream of the treatment. This is the analytical warrant for the mediator treatment.

The slip-definition and estimation literature sharpens the point. The lugged-wheel slip-ratio work (Ding et al. [\[42\]](#ref-ding2009)) provides multiple interconvertible definitions of slip ratio for a lugged wheel and two estimation methods (visual lug-trace analysis and a terramechanics-based solution), validated on single-wheel experiments. The drawbar-pull estimation via a built-in force-sensor-array wheel (Nagatani et al. [\[90\]](#ref-nagatani2009)) shows that conventional terramechanics methods misestimate the normal-stress distribution for small wheels and that direct measurement is needed for accurate drawbar pull. The stress-distribution and multi-physics interaction models (Ding et al. [\[43\]](#ref-ding2009b); Ding et al. [\[44\]](#ref-ding2014); Jiao et al. [\[66\]](#ref-jiao2010)) extend the modeling to lug effects, slip sinkage, wheel dimension, and load. The machine-learning slip-estimation work (Kruger, Rogg, and Gonzalez [\[69\]](#ref-kruger2019); Zhang et al. [\[128\]](#ref-zhang2022) on the Zhurong rover; Yang [\[56\]](#ref-h2022)) and the inverse-terramechanics surveys (Lopez-Arreguin [\[4\]](#ref-ajr2021); Lopez Arreguin [\[1\]](#ref-a2021)) document the more recent learning-based estimation of slip and regolith strength. The method ranges from sensor-instrumented experiment to data-driven estimation; the convergent finding is that slip and sinkage are estimated quantities produced during a drive. The interpretation is decisive for the dissertation: every one of these sources treats slip as an outcome to be estimated, none treats it as an exogenous input, and that universal structure is the strongest possible external support for keeping realized slip off the right-hand side of the productivity regression.

The discrete-element and gravity-effect literature supplies the gravity-scaling and high-fidelity-modeling basis ([\[35\]](#ref-collective2020); the cellular-automata extended model of Watanabe [\[120\]](#ref-y2023); the high-slip sinkage prediction of Wang [\[126\]](#ref-z2023); the various Yeo wheel-dynamics observers [\[123\]](#ref-yeo2025)). The volumetric and contact-modeling tradition (Petersen and McPhee [\[96\]](#ref-petersen2013); Petersen and McPhee [\[97\]](#ref-petersen2015)) provides closed-form contact forces for fast multibody simulation validated against drawbar-pull tests on the Canadian Space Agency Juno rover. The locked-wheel and normal-stress-distribution measurements (Fujiwara, Oshima, and Iizuka [\[47\]](#ref-fujiwara2020)) and the FPGA-acceleration work on visual odometry (Lentaris et al. [\[73\]](#ref-lentaris2015)) round out the method space. The latter bears directly on the argument because it documents the computational cost that constrains how much perception a flight processor can afford during a drive, tying the terramechanics layer back to the autonomy-timing question of Section 3.3.

The hardware-platform and suspension literature supplies the mechanical covariates. The rocker-bogie suspension analysis (Cosenza et al. [\[38\]](#ref-cosenza2023)) and the active-suspension resilient-mobility work (Ricano et al. [\[99\]](#ref-ricano2025)) characterize how suspension geometry distributes traction and rolling resistance across wheels, which is the physical basis for the actuator-class and mass-class covariates. The soft-soil parameter-uncertainty work (Gallina, Krenn, and Schäfer [\[49\]](#ref-gallina2016)) and the contact-analysis modeling (Mahon [\[78\]](#ref-mahon2016)) document how uncertainty in soil parameters propagates into mobility-simulation predictions, a caution the dissertation carries as measurement error in the constructed terrain covariates. The inflatable-rover testbed results (Apostolopoulos [\[16\]](#ref-apostolopoulos2018)) and the manned-lunar-rover wheel design (Xiao [\[118\]](#ref-xiao2016)) bound the hardware design space at its extremes. The combined-EDL-mobility planning work (Kuwata [\[119\]](#ref-y2011)) and the high-accuracy speckle-velocimetry odometry concept (Charrett, Waugh, and Tatam [\[31\]](#ref-charrett2009)) document adjacent measurement and planning approaches.

The following synthesis table organizes the terramechanics evidence by its function in the dissertation's design.

| Function | Representative source(s) | Method | What it supplies | Methodological consequence |
|---|---|---|---|---|
| Slip as a modeled outcome | Ishigami et al. [\[61\]](#ref-ishigami2006), [\[63\]](#ref-ishigami2007); Ding et al. [\[42\]](#ref-ding2009), [\[44\]](#ref-ding2014) | Semi-empirical + single-wheel validation | Slip/sinkage as functions of commanded motion, soil, slope | Justifies slip-as-mediator (bad-controls discipline) |
| Slip estimation (incl. ML) | Nagatani et al. [\[90\]](#ref-nagatani2009); Kruger et al. [\[69\]](#ref-kruger2019); Zhang et al. [\[128\]](#ref-zhang2022) | Sensor instrumentation; data-driven | Slip is an estimated, drive-produced quantity | Reinforces keeping realized slip off the RHS |
| Gravity and high-fidelity modeling | [\[35\]](#ref-collective2020); Petersen & McPhee [\[96\]](#ref-petersen2013), [\[97\]](#ref-petersen2015); Watanabe [\[120\]](#ref-y2023) | DEM, volumetric contact, multibody | Gravity-scaled wheel-soil forces | Supports terrain-covariate construction with stated uncertainty |
| Hardware/suspension basis | Cosenza et al. [\[38\]](#ref-cosenza2023); Ricano et al. [\[99\]](#ref-ricano2025); Gallina et al. [\[49\]](#ref-gallina2016) | Suspension modeling; uncertainty analysis | Actuator/mass-class covariate basis | Defines the hardware block to be conditioned on |
| Compute cost of perception | Lentaris et al. [\[73\]](#ref-lentaris2015) | HW/SW codesign, FPGA acceleration | Per-drive computational budget for VO | Links terramechanics to autonomy-timing constraint |

The convergence warrants a very-high-confidence claim that slip and wheel-soil interaction are drive-produced outcomes rather than exogenous inputs, the single most important methodological result the literature contributes to this dissertation. It also warrants a high-confidence claim that the hardware and terrain covariates the design requires are constructible from this literature, with the qualifier that soil-parameter uncertainty (Gallina et al. [\[49\]](#ref-gallina2016)) propagates measurement error into the terrain covariates that Chapter 5 must carry.

## 3.5 Surveys, path-planning generations, and the autonomy-generation taxonomy

This section argues that the survey and path-planning literature corroborates the dissertation's three-level autonomy-generation taxonomy as a real progression rather than a marketing convenience, and that it independently identifies the same gap this dissertation fills: the literature catalogs algorithmic generations and surveys mobility advances without ever partitioning fleet-level productivity between the software and hardware channels. Multiple independent surveys describe the same generational progression, and the path-planning-algorithm taxonomies converge on the boundaries the dissertation uses to define G1, G2, and G3.

The keystone survey is the past-present-future review of space-robot autonomy (Gao and Chien [\[50\]](#ref-gao2021)). Its method is a structured review across mission phases; its central finding is decisive for the dissertation's framing: "spacecraft today remain largely reliant on ground in the loop to assess situations and plan next actions, using pre-scripted command sequences," and "the ability of ground operators to predict the outcome of their plans seriously diminishes when platforms physically interact with planetary bodies, as has been experienced in two decades of Mars surface operations." The survey literature itself thus identifies the binding scarcity, ground-in-the-loop planning cycles, that onboard autonomy relaxes, which is the operational consequence in the dissertation's causal chain, and it identifies the physical-interaction uncertainty (slip, sinkage) that makes the terramechanics mediators of Section 3.4 unavoidable. Gao and Chien do not estimate the productivity payoff of relaxing that scarcity; they describe it. That descriptive-but-not-identified posture is the gap, stated by the most authoritative survey in the field.

The two mobility-and-path-planning reviews tagged for this chapter corroborate the generational structure ([\[36\]](#ref-collective2022); [\[37\]](#ref-collective2025)). The method in each is a literature taxonomy; the convergent finding is a progression from early reactive hazard-avoidance schemes, through grid-based and global planners (the GESTALT and Field D* lineage of Section 3.1), and on to learning-augmented planners now in development. The general mobile-robot path-planning reviews (Yang et al. [\[122\]](#ref-yang2023); the visual-odometry surveys of Yousif, Bab-Hadiashar, and Hoseinnezhad [\[124\]](#ref-yousif2015), Aqel et al. [\[17\]](#ref-aqel2016), and Fraundorfer and Scaramuzza [\[46\]](#ref-fraundorfer2012)) place the planetary-rover work in the broader robotics context and document the maturation of visual odometry from a research capability into an established navigation building block. The autonomy-generation construct is therefore externally corroborated: independent reviewers, using independent taxonomies, draw the same generational boundaries the dissertation uses, which raises confidence that G1, G2, and G3 are distinct generations rather than relabelings, and which the TechPort robustness check is designed to confirm with a mission-external maturity record.

The contemporary and emerging path-planning literature documents the learning-augmented frontier that the Mokyr lens reads as evidence of a still-widening propositional base. The machine-learning and learning-based planners (Yu, Wang, and Zhang 2021 on safety-constrained end-to-end lunar planning; Wang et al. [\[115\]](#ref-wang2025) on few-shot tiered planning for Mars; Chatterjee, Mitra, and Bhowmik [\[32\]](#ref-chatterjee2024) on ML obstacle avoidance), the classical and graph-based planners (Katiyar and Dutta [\[68\]](#ref-katiyar2019) on RRT in configuration space; Tao et al. [\[107\]](#ref-tao2022) on visibility-graph methods; Li et al. [\[76\]](#ref-li2015) on traversability-based avoidance; Wei, Sun, and Tian 2025 on kinematic-constrained terrain-aware planning), the battery-and-health-aware planners (Salinas-Camus, Kulkarni, and Orchard [\[103\]](#ref-salinascamus2023); the slip-rate-dependent traversability model of Sakayori [\[48\]](#ref-g2024b)), and the lunar and cooperative extensions (Chen et al. [\[33\]](#ref-chen2025) on South Pole lunar planning; Cardone [\[24\]](#ref-cardone2025) on cooperative navigation; the ExoMars-lineage onboard planning of Rusu et al. [\[100\]](#ref-rusu2013) and the path-planning thesis of Rusu) collectively document an active research frontier. The university-rover and prototype work (Zaman et al. [\[127\]](#ref-zaman2022) on the Phoenix human-assistance rover) and the general autonomous-navigation references (Autonomous navigation in cluttered environments, 2016; the Toupet et al. NTRS Enhanced Navigation record) round out the survey base. The method across these is algorithm development with simulation or analog validation; the finding is a vigorous, still-expanding set of planning techniques. The interpretation, qualified as analog and developmental evidence rather than flight evidence, is that the autonomy layer sits on a research frontier whose propositional base continues to widen. That is precisely Mokyr's signature of an extensible technique whose marginal returns can stay high, in contrast to the bounded refinement of a wheel.

The following synthesis table organizes the survey and taxonomy evidence.

| Theme | Representative source(s) | Method | Key finding | Bearing on the gap |
|---|---|---|---|---|
| Space-autonomy state of practice | Gao and Chien [\[50\]](#ref-gao2021) | Structured review across mission phases | Ground-in-the-loop is the binding scarcity; autonomy relaxes it | States the gap: payoff described, never identified |
| Mobility/path-planning generations | [\[36\]](#ref-collective2022); [\[37\]](#ref-collective2025) | Algorithm taxonomy | Reactive -> global -> learning-augmented progression | Corroborates the G1/G2/G3 construct |
| Visual-odometry maturation | Yousif et al. [\[124\]](#ref-yousif2015); Aqel et al. [\[17\]](#ref-aqel2016); Fraundorfer & Scaramuzza [\[46\]](#ref-fraundorfer2012) | VO surveys | VO matured into an established navigation block | Validates the perception substrate of the generations |
| Learning-augmented frontier | Yu et al. [\[125\]](#ref-yu2021); Wang et al. [\[115\]](#ref-wang2025); Sakayori [\[48\]](#ref-g2024b) | ML planner development + simulation | Active, expanding learning-based planning | Mokyr signature: widening propositional base |
| Classical/graph planners | Katiyar & Dutta 2019; Tao et al. [\[107\]](#ref-tao2022); Wei et al. [\[117\]](#ref-wei2025) | Algorithm development | Diverse mature planning methods | Breadth of the prescriptive technique space |
The convergence warrants a high-confidence claim that the autonomy-generation taxonomy is externally corroborated and a moderate-confidence claim, qualified by the developmental nature of much of this evidence, that the autonomy layer rests on a widening propositional base in Mokyr's sense. The field's most authoritative survey (Gao and Chien [\[50\]](#ref-gao2021)) states the dissertation's gap in its own words: the productivity payoff of relaxing the ground-in-the-loop scarcity is described across two decades of Mars operations but never identified.

## 3.6 The Angrist-Pischke and Mokyr lenses applied to the domain literature

Having surveyed the domain literature thematically, this section interprets it through the two methodological anchors. The literature review's purpose is not only to catalog what is known but to show why the known evidence cannot answer the causal question without a new design. The claim is that the domain literature is, in design-based-econometric terms, a large collection of rich descriptions with no identifying comparison, and that the Mokyr framework explains why the two channels should differ in their returns and therefore why partitioning them is worth the effort.

The Angrist-Pischke reading is developed in the methodological anchor sources (Angrist and Pischke [\[7\]](#ref-angrist2009); Angrist and Pischke [\[8\]](#ref-angrist2014); the credibility-revolution statement of Angrist and Pischke [\[10\]](#ref-angristpischke2010); the LATE and IV-identification foundations of Imbens and Angrist [\[60\]](#ref-imbensangrist1994) and Angrist, Imbens, and Rubin [\[9\]](#ref-angristimbensrubin1996)). The credibility revolution's central lesson is that empirical credibility came from shifting away from elaborate functional-form assumptions toward research designs that exploit natural or quasi-experimental variation (Angrist and Pischke [\[10\]](#ref-angristpischke2010)). Read against the domain literature, the diagnosis is exact: every domain source establishes that a rover did something and that productivity rose, but none constructs a comparison in which the autonomy generation varies while terrain and hardware are held fixed. The directed-versus-autonomous mode-choice documentation (Biesiadecki, Leger, and Maimone [\[20\]](#ref-biesiadecki2007)) comes closest, because it records a within-machine mode choice, but it stops at description and does not condition on terrain. The dissertation imports the whole logical apparatus of the anchor: the named counterfactual (a drive that could have been executed blind but was executed autonomously, or vice versa, on the same machine and the same terrain class), the within-unit comparison that holds hardware fixed, the strict bad-controls rule that keeps realized slip on the outcome side (reinforced independently by the terramechanics literature of Section 3.4), and the few-cluster-aware inference required by a fleet of three full panel members. The few-cluster and heterogeneous-effects literature (de Chaisemartin and D'Haultfoeuille [\[28\]](#ref-chaisemartin2019), 2020, 2020b; Goodman-Bacon [\[53\]](#ref-goodmanbacon2018), 2021; Callaway and Sant'Anna [\[23\]](#ref-callawaysantanna2021); the sensitivity-analysis caution of Leamer [\[72\]](#ref-leamer2010)) supplies the inferential machinery. Its relevance is that with treatment varying across generations adopted at different times, naive two-way fixed-effects coefficients can be contaminated weighted averages, so the estimator and its standard errors must be chosen with that hazard in view. Following the anchor, the dissertation protects one qualification: the between-rover generation contrast leans on the hardware controls to separate collinear channels and is therefore weaker than the within-rover autonomous-fraction contrast, a point made explicit so the contribution cannot be oversold.

The Mokyr reading is developed in the framework anchor sources (Mokyr [\[85\]](#ref-mokyr2002); the political-economy and twenty-five-centuries treatments of Mokyr [\[84\]](#ref-mokyr1998) and Mokyr [\[86\]](#ref-mokyr2013); the propositional-knowledge epistemology of Villoro [\[113\]](#ref-villoro1998); the trade-and-knowledge growth models of [\[92\]](#ref-orourke2007) [\[93\]](#ref-orourke2008); the knowledge-diffusion endogenous-growth bridge of Braunerhjelm et al. [\[21\]](#ref-braunerhjelm2010)). The central distinction is between propositional knowledge (the understanding of why something works) and prescriptive knowledge (the technique that does the work), with sustained progress requiring the prescriptive technique to rest on a widening propositional base. Read against the domain literature, the two channels fall cleanly on either side of this distinction. The mechanical platform is prescriptive improvement of a finished kind: a larger wheel, a stronger actuator, a heavier chassis are refinements of an established technique whose physics is fully understood (the terramechanics literature of Section 3.4 is the codification of that understanding) and whose marginal returns are therefore bounded. The autonomy-software generation, by contrast, rests on a propositional base in perception and planning that the survey literature of Section 3.5 shows is still widening, and it is extensible in a way a wheel is not: the same flight processor running ENav can in principle run a better planner uploaded after landing, whereas a wheel cannot be enlarged after launch. The Mokyr framework also warns that such gains are fragile and depend on a low access cost between research and deployment, which maps onto the dissertation's observation that autonomy advances reach the surface through flight-software updates while hardware is frozen at launch. Mokyr supplies the why behind the dissertation's central conjecture: the two channels should differ in their returns because one is a finished technique and the other sits on a research frontier, so a finding that the autonomy channel dominates would be not a statistical curiosity but an instance of a well-established historical pattern.

The two lenses are complementary rather than redundant, and the literature review demonstrates the complementarity concretely. Angrist and Pischke tell us that the domain literature cannot answer the causal question because it lacks an identifying comparison, and they tell us how to build one. Mokyr tells us why the answer matters and what it would imply for the technology's future trajectory, which a purely statistical result could not. A regression coefficient showing the autonomy channel dominates would be inert without Mokyr's account of why software and steel differ in their returns; Mokyr's historical claim would be merely suggestive without a credible estimate that holds terrain and hardware fixed. The dissertation needs both, and the literature surveyed here supplies the domain substance that each lens operates on.

## 3.7 Synthesis, the explicit gap, and the propositions that follow

The chapter's contribution to the dissertation's argument can now be stated compactly, because the evidence for each part has been developed thematically above. The productivity rise is genuine across the fleet, and the autonomy and hardware channels are described together and never separated (Crisp et al. [\[39\]](#ref-crisp2003); Grotzinger et al. [\[55\]](#ref-grotzinger2012); Vasavada et al. [\[110\]](#ref-vasavada2014); Farley et al. [\[45\]](#ref-farley2023); Gao and Chien [\[50\]](#ref-gao2021); Verma et al. [\[112\]](#ref-verma2025)). Separating them is decision-relevant because sols are the binding scarcity for sampling, instrument placement, and caching, and the survey literature names ground-in-the-loop planning as the constraint that onboard autonomy relaxes (Farley et al. [\[45\]](#ref-farley2023); Golombek et al. [\[52\]](#ref-golombek2014); Genova et al. [\[51\]](#ref-genova2013); Gao and Chien [\[50\]](#ref-gao2021)). The design reaches the causal mechanism through a within-rover autonomous-fraction contrast that holds hardware fixed and isolates the software channel, with the bad-controls rule, independently mandated by the terramechanics treatment of slip as a drive-produced outcome, keeping realized slip on the outcome side (Angrist and Pischke [\[7\]](#ref-angrist2009); Maimone, Cheng, and Matthies [\[79\]](#ref-maimone2007); Biesiadecki, Leger, and Maimone [\[20\]](#ref-biesiadecki2007); Helmick et al. [\[59\]](#ref-helmick2004); Ishigami et al. [\[61\]](#ref-ishigami2006)). That contrast improves on its rivals, the combined-narrative description and the naive between-rover regression that cannot separate collinear channels, with the heterogeneous-effects literature establishing why the naive estimator is hazardous (Angrist and Pischke [\[10\]](#ref-angristpischke2010); de Chaisemartin and D'Haultfoeuille [\[29\]](#ref-chaisemartin2020); Goodman-Bacon [\[54\]](#ref-goodmanbacon2021); Callaway and Sant'Anna [\[23\]](#ref-callawaysantanna2021)). The remaining risk is bounded: few-cluster inference via wild-cluster bootstrap and an explicit threat-to-validity treatment make the identification risk transparent (de Chaisemartin and D'Haultfoeuille [\[28\]](#ref-chaisemartin2019); Goodman-Bacon [\[53\]](#ref-goodmanbacon2018); Imbens and Angrist [\[60\]](#ref-imbensangrist1994); Leamer [\[72\]](#ref-leamer2010)).

The named causal mechanism the literature supports runs as follows, with each link evidenced rather than asserted. The driver is the advance in autonomy-software generation from G1 through G3, culminating in Enhanced Navigation that processes imagery and plans while driving (Verma et al. [\[112\]](#ref-verma2025); Abcouwer et al. [\[3\]](#ref-abcouwer2021)). The mechanism is that a larger share of each drive is executed under onboard hazard detection and real-time path planning rather than blind commanded motion, a mode choice documented as terrain-contingent and within-machine (Biesiadecki, Leger, and Maimone [\[20\]](#ref-biesiadecki2007); Maimone, Cheng, and Matthies [\[79\]](#ref-maimone2007)). The observable effect is a higher autonomous-drive fraction and more localized meters per sol, concentrated in hazard-dense terrain, the pattern the unconditioned Perseverance records display (Verma et al. [\[111\]](#ref-verma2023)). The operational consequence is fewer sols consumed per meter, freeing sols for sampling and caching in the daily planning cycle (Gao and Chien [\[50\]](#ref-gao2021); Farley et al. [\[45\]](#ref-farley2023)). The strategic implication is that surface mobility productivity may be bought more cheaply with software than with mass, with the cost paid in development rather than at launch and the capability uploadable to a rover already on the surface, an implication the Mokyr lens grounds in the extensibility of a knowledge-intensive technique (Mokyr [\[85\]](#ref-mokyr2002)). Where the chain rests on unconditioned aggregates rather than terrain-controlled estimates, the confidence is held at moderate and the gap-closing estimation is named as the work that would raise it; where the chain rests on measured flight-engineering results, the confidence is high.

The explicit gap statement is this. The domain literature establishes, with high-grade flight evidence, that each rover was both a better machine and a smarter one, and that per-sol productivity rose. It supplies the dependent variable (verified localized path lengths from the PDS-backed localization analyses of Section 3.3), the treatment construct (the externally corroborated G1/G2/G3 taxonomy of Sections 3.1 through 3.3 and 3.5), the hardware and terrain covariates (the terramechanics and suspension literature of Section 3.4), and the within-machine variation in autonomous fraction that the identification requires (the mode-choice documentation of Section 3.1). What it does not supply, anywhere, is a comparison that holds terrain and hardware fixed and partitions the productivity gain between the autonomy and mechanical channels. The collinearity of the two channels (upgraded together on every mission) is acknowledged implicitly by every source and confronted explicitly by none. The most authoritative survey states the gap in its own words: the payoff of relaxing the ground-in-the-loop scarcity is described across two decades of operations and never identified (Gao and Chien [\[50\]](#ref-gao2021)). This is the gap, and it is not a gap of missing data but a gap of missing design.

From this gap the dissertation's propositions follow directly, and they restate the prospectus hypotheses without paraphrasing their operative clauses. The null proposition, H0, is that conditional on terrain class, cross-rover and within-rover variation in per-sol mobility productivity is explained by mechanical platform covariates (wheel diameter, actuator and mass class, available drive energy) and not by autonomy-software generation, with the autonomy-generation coefficients jointly indistinguishable from zero once hardware covariates are included. The contribution proposition, H1, is that conditional on terrain class and on the same hardware covariates, autonomy-software generation and onboard hazard-detection capability explain the larger share of the variation in per-sol mobility productivity, with the autonomy-generation coefficients jointly significant and remaining so after hardware covariates are included, and accounting for more of the between-rover and within-rover variance than the hardware block does. Three subsidiary propositions follow from the mechanism and are testable independently. First, the within-rover autonomous-drive-fraction coefficient is expected to be positive and to survive rover and terrain-class fixed effects, because within a single machine the hardware is constant and only the software's contribution varies. Second, the autonomy effect is expected to be largest in rougher, hazard-dense terrain classes, where onboard hazard detection most reduces the blind-driving penalty, and smallest on benign flats where blind commanded driving is already near-optimal. Third, the autonomy-generation block is expected to retain joint significance and explanatory share after the hardware block enters; a collapse of the block when hardware enters would reject H1 and let H0 stand. Each proposition is falsifiable in the strong sense that the design specifies in advance what evidence would reject it (Angrist and Pischke [\[7\]](#ref-angrist2009)), and none is a finding: all are design-stage expectations to be tested on the assembled panel, labeled as such.

The chapter closes where it opened. The literature is rich in description and thin in identification. It has done the indispensable work of documenting four flight rovers, three autonomy generations, the terramechanical limits of the platform, and the algorithmic frontier of the software, and in doing so it has assembled, piece by piece, every ingredient the gap-closing design requires except the design itself. The next chapter takes those pieces, the PDS localization products, the NTRS performance reports, and the TechPort maturity records, and shows how each is operationalized into the measured quantities the propositions above will be tested against, while keeping the drive-produced mediators that the terramechanics literature so clearly identifies firmly on the outcome side where the bad-controls discipline requires them to stay.



# Chapter 4: Data and Measurement

## 4.0 Chapter thesis and orientation

The dependent variable, the treatment, the hardware covariates, and the terrain fixed effect that the dissertation requires are all constructible from three public, mission-external archives, and the measurement design described in this chapter is what converts the autonomy-versus-hardware question into estimable quantities while keeping every post-treatment mediator off the right-hand side. That is the chapter's thesis, and it is a stronger claim than it may first appear. It is not merely that data exist. Data of some kind always exist for a flown mission. The claim is that the specific quantities the identification strategy of Chapter 5 demands, a localized per-sol traverse distance, a per-drive share of distance executed under onboard autonomy, an a priori terrain class assigned before the drive rather than inferred from its outcome, and a hardware-covariate block that is fixed at the level of the individual machine, can each be operationalized from a named, citable, access-documented source, and that the one quantity for which a clean published series is not guaranteed, the per-drive autonomous-drive fraction, can be reconstructed by a procedure stated here in advance and whose measurement error is carried forward honestly rather than hidden. Measurement is where this dissertation either earns its identification or forfeits it, and the chapter treats measurement as a first-order design problem, not as a preliminary to the real work.

The problem this chapter addresses can be set out in four steps. The current state is that the productivity history of the Mars rover fleet is documented in scattered engineering reports, mission overviews, and a public planetary-data archive, but it has never been assembled into a single panel with a consistent unit of analysis and a consistent set of constructs. The desired state is a drive-sol panel in which each row carries a localized distance, an elapsed-sol count, an autonomy-generation label, an autonomous-drive fraction, a hardware-covariate vector, and a terrain class, every field traceable to a named source and a documented access path. The gap is that no such panel has been published, so the constructs must be defined and the sources must be mapped to them explicitly, with the seams between sources made visible rather than papered over. The consequence of leaving the gap unaddressed is that any estimate of the autonomy effect would rest on undocumented measurement choices, and an estimate whose measurement is undocumented cannot be falsified, because a critic cannot tell whether a null result reflects the absence of an effect or the corruption of the measure. This chapter closes that gap by specifying the measurement design completely, the precondition for the falsifiable contribution stated in the prospectus.

A note on confidence and on design-stage honesty is owed at the outset, because the integrity of the whole dissertation depends on it. This is a design-stage document. The panel described here has not been assembled, and no field in it has been populated with a value drawn from the archives. Where the chapter states what a measure will be, it is describing a construction procedure, not reporting a number. Where the chapter gives an illustrative figure to make a construct concrete, the figure is labeled illustrative and is not an estimate. The measurement claims carry calibrated confidence: the construction of the dependent variable from localized PDS traverse products is rated high confidence because the products and their access paths are documented in the mission and localization literature; the construction of the autonomous-drive fraction is rated moderate confidence because the public record does not guarantee a clean per-drive series and a reconstruction procedure must do part of the work; the construction of the terrain class is rated moderate confidence because the class is built by the candidate from orbital and rover-derived inputs rather than read off a published label. These gradings are stated where each construct is defined, and the evidence that would raise or lower each one is named. Every major measurement claim in this chapter is built the same disciplined way: it rests on cited sources, it states the reasoning that connects each source to the construct it supports, it qualifies the strength of that inference, and it names the threat the design must answer if the inference is to hold.

The chapter proceeds dataset by dataset. Section 4.1 treats the Planetary Data System traverse and localization archives, the source of the dependent variable. Section 4.2 treats the NASA Technical Reports Server performance reports, the source of the autonomy-generation classification and the autonomous-drive fraction. Section 4.3 treats the NASA TechPort technology-readiness records, the independent maturity classification used to keep the generation labels honest. Section 4.4 presents the single integrated operationalization table, construct by construct, giving for each variable its operational definition, its source, and its scale. Section 4.5 develops the terramechanics-based construction of the hardware and terrain covariates, because those covariates are not read directly from a single file but are built from a substantial modeling literature. Section 4.6 treats data quality, validation against known values, and the access and ethics of the public archives. The order is deliberate: the reader meets each source before meeting the table that depends on it, and meets the table before meeting the quality checks the table must survive.

## 4.1 The PDS rover traverse and localization archives

The first dataset is the Planetary Data System Geosciences Node archive of rover localization and traverse products, and this section argues that these archives are the correct and sufficient source for the dependent variable, mobility productivity, for the three full panel members. The argument rests on the mission and localization literature that documents what these products contain and how the surface position they record is tied to orbital basemaps. Productivity, operationalized as localized meters per sol or its inverse, is a function only of distance driven and sols elapsed, both of which the traverse products resolve at the drive-sol level, and the body of work on rover localization within orbital maps establishes the accuracy and the method by which surface traverse is registered to a common spatial frame. The sufficiency holds for the Mars Exploration Rovers, Curiosity, and Perseverance, and not for Sojourner, whose record predates the traverse-product standard. The one threat the design must answer is that localization error is itself a source of measurement error in the distance variable, and that error is addressed in Section 4.6.

### 4.1.1 Provenance and access

The Planetary Data System is NASA's permanent archive for planetary mission data, and its Geosciences Node is the node responsible for surface-mission geoscience and mobility products. The localization and traverse products for the three flight rovers are deposited there as part of each mission's archiving obligation, and they are accessible through the node's public archive interface without restriction or fee. The provenance chain is therefore short and auditable: the rover's onboard and ground-reconstructed position estimates are processed by the mission localization team, registered to orbital basemaps, and deposited as archive products that any researcher can retrieve. The dissertation treats the access path itself as a documented artifact, recorded in the data-provenance appendix, so that the assembled panel is reproducible from public sources by a third party.

The registration of surface position to orbital basemaps is the methodological hinge of these products, and it is documented in the localization literature. The matching of ground-level rover imagery to orbital imagery, the technique that ties a rover's local traverse to a planet-fixed frame, is established for the Mars rovers (Di, Liu, and Yue [\[40\]](#ref-di2011)), and the long-range localization of a rover by matching its onboard sensing to orbital elevation maps is demonstrated in the field-robotics literature (Carle, Furgale, and Barfoot [\[25\]](#ref-carle2010)). For Perseverance specifically, the rover's path has been estimated through visual odometry and registered against the mission frame in independent analyses (Andolfo, Petricca, and Genova [\[6\]](#ref-andolfo2022); Andolfo [\[101\]](#ref-s2021); Andolfo [\[102\]](#ref-s2022)), and a stereo-vision-based pose-estimation pipeline ties the rover's position to the orbital basemap at high precision (Di et al. [\[41\]](#ref-di2022)). The interpretation for this dissertation is that the localized path length, the quantity the dependent variable is built from, is not a raw odometric guess but a product registered to a common spatial frame by a documented method, which is what makes per-sol distances comparable across rovers that landed at different sites.

### 4.1.2 Coverage
Coverage of the archive spans the full surface traverse record of the three flight rovers. For the Mars Exploration Rovers, the record runs from the 2004 landings through the multi-year extended missions, encompassing the kilometers-scale traverses documented in the mission overviews (Crisp et al. [\[39\]](#ref-crisp2003)). For Curiosity, coverage runs from Bradbury Landing through the Gale crater traverse, the record described in the Mars Science Laboratory mission overviews (Grotzinger et al. [\[55\]](#ref-grotzinger2012); Vasavada et al. [\[110\]](#ref-vasavada2014)). For Perseverance, coverage runs from the 2021 Jezero landing through the first science campaign on the crater floor and beyond (Farley et al. [\[45\]](#ref-farley2023)). The temporal span is therefore roughly two decades, and the cumulative distance across the fleet is on the order of tens of kilometers per long-lived rover. At the drive-sol unit of analysis this yields a panel of thousands of rows, even though the cross-section of rovers is only three. This asymmetry, many drive-sols nested within few rovers, is the defining structural feature of the panel, and it is why the inference plan of Chapter 5 leans on within-rover variation and on few-cluster-aware standard errors.

### 4.1.3 Unit of analysis

The primary unit of analysis is the drive-sol: one rover, one sol, one or more commanded drives aggregated to the sol. This choice is consequential, and it is defended here. A sol, the Martian solar day, is the natural unit of the surface operations cycle, because the rover is commanded once per sol on the basis of the previous sol's data (Crisp et al. [\[39\]](#ref-crisp2003)). The sol is therefore the period over which a planner budgets distance and over which productivity is naturally measured. Aggregating multiple commanded drives within a sol to a single drive-sol row matches the unit of analysis to the unit of decision, the sol-level plan. The panel is thus an unbalanced panel of drive-sols nested within rovers. A secondary unit, the individual commanded drive, is used for robustness where the archive resolves multiple drives within a sol. The drive is the level at which the autonomy mode (autonomous versus blind commanded) is actually selected, and the drive-level unit lets the within-rover autonomous-fraction contrast be estimated at its native resolution. The two units are complementary: the drive-sol matches the planning cycle and serves as the primary unit for the productivity outcome, while the drive matches the autonomy decision and serves as the finer unit for the treatment.

### 4.1.4 Known biases and limitations

Three biases in this source are stated plainly. First, Sojourner predates the PDS traverse-product standard. The 1997 microrover's traverse record does not exist in the archive at the drive-sol resolution the panel requires, and the prospectus already designates Sojourner a boundary case rather than a fourth full panel member. The dissertation treats Sojourner qualitatively, as the zero point of the autonomy trajectory, and does not promise a populated Sojourner row series. This is a coverage limitation, not a defect to be filled by fabrication, and the chapter is explicit that no Sojourner drive-sol data are manufactured to close it. Second, localization error propagates into the distance measure. Because the localized path length is a registered estimate, its error budget is the localization pipeline's error budget, and a systematic localization bias on one rover would bias that rover's distances relative to the others. Section 4.6 specifies the validation against known mission-reported cumulative distances that bounds this error. Third, the traverse products record where the rover went, not why it went there, so terrain along the path is endogenous to the rover's route choice. The terrain-class fixed effect and the within-terrain-class comparison of Chapter 5 are designed to address this selection but cannot fully dissolve it. These three biases are the price of using a real flown archive rather than a designed experiment, and naming them is part of the design's honesty.

## 4.2 The NTRS AutoNav and Enhanced Navigation performance reports

The second dataset is the NASA Technical Reports Server collection of engineering reports that characterize autonomous-navigation performance by generation. This section argues that these reports are the correct source for the treatment variable, autonomy-software generation, and for the continuous autonomous-drive fraction that is the dissertation's preferred treatment measure. The primary engineering reports describe each generation's autonomy stack and its behavior in flight. The autonomy generation and the autonomous-drive fraction are properties of the navigation software's design and operation, which these reports document directly, and not properties that must be inferred from the productivity outcome. Independent reports across the three generations converge onto a consistent account of what each generation added. One caveat must be carried forward without softening: the reports document autonomy capability and behavior cleanly but do not guarantee a ready-made per-drive autonomous-fraction series, so the fraction must be reconstructed. The threat the design must answer is measurement error in that reconstruction, and the section states the reconstruction procedure precisely so the error is bounded and transparent.

### 4.2.1 Provenance and access

The NASA Technical Reports Server is the agency's public index of technical reports, conference papers, and journal articles produced by NASA centers and their contractors, and it is accessible without restriction. The reports relevant here are the engineering accounts of the navigation systems of each rover generation. For the Mars Exploration Rover generation, the load-bearing report is the two-year visual-odometry account (Maimone, Cheng, and Matthies [\[79\]](#ref-maimone2007)), supplemented by the directed-versus-autonomous-driving tradeoff analysis (Biesiadecki, Leger, and Maimone [\[20\]](#ref-biesiadecki2007)), the visual-odometry conference treatment (Cheng, Maimone, and Matthies [\[34\]](#ref-cheng2006)), and the high-slip path-following account (Helmick et al. [\[59\]](#ref-helmick2004)). For the Mars 2020 generation, the definitive report is the Enhanced Navigation description (Verma et al. [\[112\]](#ref-verma2025)). These reports are the source of both the categorical generation classification and the behavioral measures from which the autonomous-drive fraction is reconstructed. The provenance is again short and auditable: each report is a primary engineering account by the teams that built or operated the system, indexed in a public archive.

### 4.2.2 Operationalizing the autonomy-software generation

The treatment variable is autonomy-software generation, a categorical indicator with three levels fixed by the dissertation's shared specification: G1, the Mars Exploration Rover class AutoNav with visual odometry; G2, the Mars Science Laboratory inherited and extended stack; and G3, the Mars 2020 Enhanced Navigation. The operationalization of these levels rests on the documented content of each generation's stack rather than on the mission's marketing of it. G1 is identified by the conjunction of onboard stereo hazard assessment, visual odometry for slip-aware position estimation, and global path planning, all documented in the MER engineering reports (Maimone, Cheng, and Matthies [\[79\]](#ref-maimone2007); Biesiadecki, Leger, and Maimone [\[20\]](#ref-biesiadecki2007); Helmick et al. [\[59\]](#ref-helmick2004)). G2 is identified as the inheritance and extension of that stack onto the larger Curiosity platform, documented in the Mars Science Laboratory mission overviews (Grotzinger et al. [\[55\]](#ref-grotzinger2012); Vasavada et al. [\[110\]](#ref-vasavada2014)); the G2 label captures a stack recognizably descended from G1 but operating on a heavier machine in the harsher wheel-wear environment of Gale crater (Arvidson et al. [\[18\]](#ref-arvidson2017)). G3 is identified by the specific change that Enhanced Navigation introduced, the processing of imagery and planning of path while the rover is still in motion, documented in the Perseverance navigation report (Verma et al. [\[112\]](#ref-verma2025)). The construct is therefore a coarse three-level ordering of distinct, documented software capabilities, and its coarseness is acknowledged: the continuous autonomous-drive fraction, defined next, is the finer and preferred measure, and the TechPort records of Section 4.3 are the external check that the three levels are genuinely distinct generations rather than relabelings.

### 4.2.3 Operationalizing the autonomous-drive fraction

The autonomous-drive fraction is the share of a drive's distance executed under onboard autonomous navigation rather than blind commanded motion. It is the dissertation's preferred treatment measure because it varies within rover and therefore identifies the autonomy effect with hardware held constant at the level of the individual machine. The measurement of this fraction is the single highest-value and highest-difficulty step in the whole data design, and the chapter is explicit about both its importance and its difficulty. The basis for the construct is the engineering reports' documentation that drives are composed of segments driven in distinct modes. The MER tradeoff report states that driving strategies alternately use more or less onboard autonomy to maximize drive speed and distance at the cost of greater command complexity, and that solar energy generally permitted at most about four hours of driving per sol (Biesiadecki, Leger, and Maimone [\[20\]](#ref-biesiadecki2007)), which establishes that the autonomous share is a real, varying, segment-level property of each drive rather than a fixed rover attribute. The Enhanced Navigation report documents the corresponding G3 capability, the raised autonomous share that thinking-while-driving makes possible (Verma et al. [\[112\]](#ref-verma2025)).

The difficulty is that a clean, published, per-drive autonomous-fraction series across all generations is not guaranteed in the public record, and the prospectus and expansion plan both flag this as the moderate gap that most needs closing before execution. The dissertation does not pretend the series exists. Instead it states the reconstruction procedure. The autonomous-drive fraction for a drive is reconstructed as the ratio of distance executed under onboard autonomous navigation to total localized distance for that drive, where the autonomous distance is identified from the command and telemetry record of the drive, which distinguishes blind commanded motion from AutoNav-controlled motion. Where the per-drive command record is available in the archive, this ratio is computed directly. Where it is not available at the drive level, the fraction is reconstructed at the coarser sol level or generation level from the aggregate behavioral statistics the engineering reports provide, and the coarser reconstruction is flagged in the panel with a measurement-resolution indicator so that the analysis of Chapter 6 can test the sensitivity of the autonomy estimate to measurement resolution. The confidence in this construct is rated moderate, lower than the confidence in the dependent variable, precisely because of this reconstruction. The evidence that would raise the confidence is the recovery of a per-drive command-mode record for each generation from the archives; the evidence that would lower it is the discovery that the command record does not reliably distinguish autonomous from blind segments at the drive level for one or more generations, in which case the within-rover identification would have to rest on the coarser sol-level measure and the measurement-error caveat of Chapter 5 would bind more tightly.

### 4.2.4 The bad-controls boundary at the measurement stage

A measurement decision that looks technical is in fact the load-bearing identification decision, and it is made here rather than deferred to the design chapter because it is a decision about what each source is allowed to contribute. The autonomous-drive fraction is constructed from the commanded properties of the drive, the mode in which each segment was commanded, and not from any property realized during or after the drive. Realized slip, realized drive duration, and realized wheel-soil interaction are recorded in the archive and are scientifically interesting, but they are post-treatment mediators, produced by the drive, partly as a consequence of how the autonomy software chose to drive, and they are kept on the outcome side and never enter the treatment construction or the covariate block (Angrist and Pischke [\[7\]](#ref-angrist2009)). This is the bad-controls discipline applied at the measurement stage: a variable's eligibility to be a covariate is decided by whether it is fixed before the drive (eligible) or produced by the drive (ineligible mediator), and the chapter records this classification for every variable in the operationalization table so that no realized quantity can leak into the right-hand side by accident. The reason for this strict line is that conditioning on a post-treatment mediator absorbs part of the effect of interest and biases the estimate toward the null (Angrist and Pischke [\[7\]](#ref-angrist2009); Angrist and Pischke [\[8\]](#ref-angrist2014)), so a measurement design that allows realized slip onto the right-hand side would manufacture a false rejection of H1. The line is drawn at measurement so that it cannot be crossed at estimation.

## 4.3 The TechPort mobility-technology TRL records

The third dataset is the NASA TechPort record of the technology-readiness-level history of mobility technologies, and this section makes a narrow and specific case: these records supply an independent, mission-external classification of the maturity of each autonomy and mobility technology at the time each rover was designed, used solely as a robustness check on the autonomy-generation indicators so that the generation classification does not rest only on mission self-description. The TechPort records document the TRL history of named technologies. An external maturity timeline, built for technology-management purposes rather than for this dissertation, provides a check on whether the G1, G2, and G3 labels track genuine maturity steps, and because TRL is a NASA-standard, cross-program maturity scale, a TRL trajectory is comparable across the autonomy and mobility technologies of different missions. One point must be stated firmly: TechPort entries are program records, not peer-reviewed literature, so they are named as a dataset and cited as data provenance, not represented as graded corpus citations; writers and readers must not expect a corpus key for each TechPort record. The construct-validity worry the robustness check answers is that the three generation levels might be mission marketing rather than real technological discontinuities.

The access path is the public TechPort interface, and the records are retrieved by technology and by the mission that flew them. The use in the dissertation is confined to a crosswalk: each autonomy-generation level is mapped to the TechPort TRL history of its defining technologies (the AutoNav and visual-odometry technologies for G1, the inherited stack for G2, and the Enhanced Navigation technology for G3), and the crosswalk is reported in the supplementary appendix. If the TRL trajectories show three distinct maturity steps aligned with the three generation labels, the construct validity of the autonomy-generation variable is reinforced. If, contrary to expectation, the TRL records show no maturity discontinuity at one of the generation boundaries, the corresponding generation label is flagged as weakly supported and the robustness analysis of Chapter 6 downweights it. This is the external check that the prospectus calls for, and its presence is what lets the dissertation claim that the generation classification is grounded in independent maturity records and not only in each mission's account of itself.

## 4.4 Integrated operationalization of every variable

This section presents the single integrated operationalization table, the heart of the chapter and the formal carry-forward of the prospectus data section. The claim is that every variable named in the dissertation's fixed specification, the dependent variable in both of its parallel forms, the treatment in both of its forms, the hardware-covariate block, the terrain fixed effect and covariates, and the mediators that are excluded by construction, has a stated operational definition, a named source, and a defined scale, and that the table makes the bad-controls classification explicit for each one. The table reproduces the notation of the dissertation's fixed specification exactly, so that the constructs defined here map without ambiguity onto the estimating equation carried into Chapter 5.

For reference, the estimating equation that these variables populate, fixed verbatim across the dissertation, is, for drive-sol *i* on rover *r* in terrain class *c*:

\[
\text{Productivity}_{irc} = \beta_1 \, \text{AutonomyGen}_{r}
                 + \beta_2 \, \text{Hardware}_{r}
                 + \gamma \, \text{Terrain}_{ic}
                 + \alpha_{r} \quad \text{(rover fixed effect)}
                 + \delta_{c} \quad \text{(terrain-class fixed effect)}
                 + \epsilon_{irc} \qquad\qquad (1)
\]

The operationalization table below defines each term of this equation as a measured construct.

| Construct (role in the model) | Operational definition | Source | Scale | Bad-controls status |
|---|---|---|---|---|
| Mobility productivity, meters-per-sol (dependent, primary) | Localized path length driven on a drive-sol divided by elapsed sols for that drive-sol; distance is the registered traverse-product path length, not straight-line displacement | PDS Geosciences Node traverse and localization products (Di, Liu, Yue [\[40\]](#ref-di2011); Carle, Furgale, Barfoot [\[25\]](#ref-carle2010); Di et al. [\[41\]](#ref-di2022)) | Continuous, meters per sol, non-negative | Outcome |
| Mobility productivity, sols-per-meter (dependent, parallel) | Inverse of meters-per-sol; the quantity a mission planner budgets | Same as above | Continuous, sols per meter, positive | Outcome |
| AutonomyGen, generation (treatment, categorical) | Three-level indicator: G1 (MER AutoNav with visual odometry), G2 (MSL inherited/extended stack), G3 (Mars 2020 Enhanced Navigation), assigned from documented stack content | NTRS engineering reports (Maimone, Cheng, Matthies [\[79\]](#ref-maimone2007); Biesiadecki, Leger, Maimone [\[20\]](#ref-biesiadecki2007); Verma et al. [\[112\]](#ref-verma2025)); TechPort TRL crosswalk as check | Ordinal, three levels | Treatment (rover-level) |
| AutonomyGen, autonomous-drive fraction (treatment, continuous, preferred) | Share of a drive's localized distance executed under onboard autonomous navigation rather than blind commanded motion, from the command/telemetry mode record | NTRS reports and the command/telemetry record (Biesiadecki, Leger, Maimone [\[20\]](#ref-biesiadecki2007); Verma et al. [\[112\]](#ref-verma2025)); reconstructed per the Section 4.2.3 procedure | Continuous, 0 to 1 | Treatment (within-rover varying) |
| Hardware, wheel diameter | Nominal drive-wheel diameter of the rover | Published mission descriptions (Crisp et al. [\[39\]](#ref-crisp2003); Grotzinger et al. [\[55\]](#ref-grotzinger2012); Maki et al. [\[80\]](#ref-maki2020)); TechPort | Continuous, meters; constant within rover | Covariate (pre-treatment) |
| Hardware, mass class | Rover mass class (microrover / MER-class / MSL-class) | Mission descriptions (Crisp et al. [\[39\]](#ref-crisp2003); Grotzinger et al. [\[55\]](#ref-grotzinger2012)) | Ordinal; constant within rover | Covariate (pre-treatment) |
| Hardware, actuator class | Drive-actuator capability class | Mission descriptions; terramechanics actuator analyses (Ishigami et al. [\[61\]](#ref-ishigami2006); Ding et al. [\[44\]](#ref-ding2014)) | Ordinal; constant within rover | Covariate (pre-treatment) |
| Hardware, nominal available drive energy per sol | Nominal energy available for driving per sol, from the power subsystem description | Mission descriptions (Crisp et al. [\[39\]](#ref-crisp2003); Biesiadecki, Leger, Maimone [\[20\]](#ref-biesiadecki2007)) | Continuous, watt-hours per sol; near-constant within rover | Covariate (pre-treatment) |
| Terrain class (fixed effect, delta_c) | A priori terrain class assigned from orbital basemaps and physical-properties characterizations before the drive | Orbital and rover-derived terrain characterizations (Golombek et al. [\[52\]](#ref-golombek2014); Arvidson et al. [\[18\]](#ref-arvidson2017)) | Categorical, class labels | Covariate / fixed effect (a priori) |
| Terrain, slope (covariate within Terrain_ic) | A priori local slope along the planned path from orbital topography | Orbital basemaps; localization registration (Di, Liu, Yue [\[40\]](#ref-di2011); Carle, Furgale, Barfoot [\[25\]](#ref-carle2010)) | Continuous, degrees | Covariate (a priori) |
| Terrain, physical-properties index (covariate within Terrain_ic) | A priori index of terrain physical properties (e.g., load-bearing strength, ripple density) from orbital and early-sol characterization | Golombek et al. [\[52\]](#ref-golombek2014); Arvidson et al. [\[18\]](#ref-arvidson2017) | Continuous index | Covariate (a priori) |
| Realized slip (mediator, EXCLUDED) | Slip ratio realized during the drive | Recorded in archive; terramechanics slip-estimation literature (Ding et al. [\[42\]](#ref-ding2009); Kruger, Rogg, Gonzalez [\[69\]](#ref-kruger2019); Zhang et al. [\[128\]](#ref-zhang2022)) | Continuous; NOT on the right-hand side | Mediator (post-treatment) |
| Realized wheel-soil interaction (mediator, EXCLUDED) | Realized wheel-soil interaction quantities (sinkage, drawbar pull) during the drive | Terramechanics literature (Ishigami et al. [\[63\]](#ref-ishigami2007); Ding et al. [\[44\]](#ref-ding2014); Nagatani et al. [\[90\]](#ref-nagatani2009)) | Continuous; NOT on the right-hand side | Mediator (post-treatment) |
The table is the formal operationalization the prospectus promised, and three of its features carry the argument. First, the dependent variable is built from localized path length, not straight-line displacement. The distinction is between measuring the path the rover actually drove and measuring its net progress. Localized length means the productivity measure rewards the rover for ground covered even when the route doubles back around hazards, which is the behavior the autonomy software produces, so the measure does not perversely penalize the autonomy channel for driving the longer-but-safer path. Second, the autonomous-drive fraction is the only treatment construct that varies within rover, and the table records that status explicitly, because the within-rover variation does the identifying work; the generation indicator, by contrast, is constant within rover and identifies the effect only between rovers. Third, the table draws the bad-controls line as a column, not a footnote: realized slip and realized wheel-soil interaction are listed in the table so that their exclusion is documented and auditable, and so that no future analyst can quietly promote a mediator to a covariate.

## 4.5 Terramechanics-based construction of the hardware and terrain covariates

The hardware-covariate block and the terrain covariates are not read directly from a single archive file; they are constructed, and the construction rests on a substantial terramechanics and wheel-soil-interaction literature. This section argues that this literature supplies a principled basis for the hardware and terrain covariates and supplies the physical theory that justifies the exclusion of realized slip and realized wheel-soil interaction as mediators. The wheel-soil-interaction models relate a wheel's geometry, the actuator's capability, and the soil's properties to the motion that results, and in doing so they tell us which quantities are properties of the machine and the terrain (eligible covariates) and which are products of the interaction between them during a drive (ineligible mediators). Independent terramechanics models, across laboratory, simulation, and analog studies, converge onto a consistent decomposition of the wheel-soil system. Most of this literature is laboratory, single-wheel, or simulation work rather than flight data, so the inference from it to flight covariates is graded as modeling evidence and the construction is validated against flight-reported values where possible. The threat the design must answer is that a terramechanics model calibrated on Earth or in simulation may misclassify a covariate, which is why the construction is anchored to mission-published hardware descriptions and not to the models alone.

The foundational terramechanics for planetary rovers establishes the wheel-soil system as the object of study and gives the traction-control analysis on which later work builds (Ishigami et al. [\[61\]](#ref-ishigami2006)), with the steering-maneuver model on loose soil extending it to turning (Ishigami et al. [\[63\]](#ref-ishigami2007)) and the path-following-with-slip-compensation work closing the control loop (Ishigami, Nagatani, Yoshida [\[62\]](#ref-ishigami2006b)). The interaction-mechanics model for rigid driving wheels on sandy terrain, which accounts for multiple physical effects at once, supplies the most complete decomposition of the wheel-soil system into machine properties and soil properties (Ding et al. [\[44\]](#ref-ding2014)), and the slip-ratio definition-and-estimation work makes explicit that slip is a realized property of a drive rather than a fixed attribute of the machine (Ding et al. [\[42\]](#ref-ding2009); Ding [\[43\]](#ref-ding2009b)). The drawbar-pull estimation work using a built-in force-sensor array (Nagatani et al. [\[90\]](#ref-nagatani2009)) and the lunar-rover wheel-terrain model for slope climbing (Jiao et al. [\[66\]](#ref-jiao2010)) round out the account of how commanded motion converts into actual progress, and the discrete-element modeling of wheel-soil interaction under varying gravity ([\[35\]](#ref-collective2020)) is the source that warrants treating gravity as fixed across the Mars panel while flagging it as the boundary on lunar generalization.

This literature does three things for the covariate construction. First, it identifies the hardware covariates that matter: wheel diameter and geometry, actuator capability, and mass are the machine-side inputs that every model takes as fixed parameters, so they form the natural hardware-covariate block, and they are read from the mission hardware descriptions (Crisp et al. [\[39\]](#ref-crisp2003); Grotzinger et al. [\[55\]](#ref-grotzinger2012); Maki et al. [\[80\]](#ref-maki2020)) with the terramechanics literature establishing why they are the right ones. Second, it identifies the terrain covariates that matter: the soil's load-bearing strength and the local slope are the terrain-side inputs the models take as fixed for a given patch, so the terrain physical-properties index and slope are the natural terrain covariates, read from the orbital and early-sol characterizations (Golombek et al. [\[52\]](#ref-golombek2014); Arvidson et al. [\[18\]](#ref-arvidson2017)). Third, and for identification this matters most, it establishes that slip and sinkage and drawbar pull are outputs of the wheel-soil interaction, computed by the models from the machine and soil inputs and the commanded motion. That physical fact justifies classifying realized slip and realized wheel-soil interaction as mediators rather than covariates (Ding et al. [\[42\]](#ref-ding2009); Ishigami et al. [\[63\]](#ref-ishigami2007)). The bad-controls rule in this dissertation is therefore not merely an econometric convention imported from Angrist and Pischke; it is underwritten by the physics of terramechanics, which says that slip is downstream of the machine, the soil, and the drive, and is therefore exactly the kind of post-treatment quantity that must stay on the outcome side.

The broader wheel-soil and slip-estimation literature is used to validate the covariate construction and to bound its error rather than to provide values directly. The current-based validation of a wheel-soil interaction model for trafficability (Liu et al. [\[77\]](#ref-liu2024)), the volumetric contact model and its experimental validation (Petersen and McPhee [\[96\]](#ref-petersen2013); Petersen and McPhee [\[97\]](#ref-petersen2015)), the high-slip-angle and skid model (Pavlov and Johnson [\[95\]](#ref-pavlov2024)) and the soil-displacement trenching analysis (Pavlov and Johnson [\[94\]](#ref-pavlov2019)), the manned-lunar-rover elastic-wheel model (Xiao [\[118\]](#ref-xiao2016)), the stress-distribution model (Ding [\[43\]](#ref-ding2009b)), the cellular-automata extended terramechanics model (Watanabe [\[120\]](#ref-y2023)), the high-slip-sinkage linear prediction (Wang [\[126\]](#ref-z2023)), the locked-wheel normal-stress measurement (Fujiwara, Oshima, Iizuka [\[47\]](#ref-fujiwara2020)), the real-time terramechanics simulation framework (Kruuse et al. [\[70\]](#ref-kruuse2024)), the scalable-bulldozing lateral-interaction model (Yeo [\[123\]](#ref-yeo2025)), the terramechanics-enhanced driving-force observer (Yeo et al. [\[123\]](#ref-yeo2025)), the slip-and-sinkage estimation on deformable terrain (Yang [\[56\]](#ref-h2022)), and the slip-rate-dependent traversability model for path planning (Sakayori [\[48\]](#ref-g2024b)) collectively establish the range of physically plausible wheel-soil behavior against which the constructed covariates are sanity-checked. The machine-learning terramechanics surveys (Lopez-Arreguin [\[4\]](#ref-ajr2021); Lopez-Arreguin [\[1\]](#ref-a2021)) and the in-situ regolith-strength inverse-terramechanics survey document the frontier of learning-based estimation, which bears on the Mokyr reading of the autonomy channel as the extensible layer but is used here only as context for the covariate construction, not as a source of flight covariates. The early wheel-soil-system-constant analyses ([\[11\]](#ref-anon1972); [\[12\]](#ref-anon1973)) and the rocker-bogie suspension study (Cosenza et al. [\[38\]](#ref-cosenza2023)) anchor the mechanical-platform side of the covariate block in the long terramechanics tradition and in the suspension geometry that converts wheel motion into chassis progress. The Zhurong slip-estimation study (Zhang et al. [\[128\]](#ref-zhang2022)) and the slip-and-sinkage estimation work (Yang [\[56\]](#ref-h2022)) extend the validation set beyond the NASA fleet, which strengthens the external check on the slip-as-mediator classification.

A second strand, the visual-odometry and pose-estimation literature, is the source that validates the distance measure and, in part, the autonomous-fraction reconstruction, because the same perception pipeline that estimates the rover's position is the pipeline whose autonomous operation the fraction measures. The two-year flight account of visual odometry on MER (Maimone, Cheng, Matthies [\[79\]](#ref-maimone2007)), the conference treatment (Cheng, Maimone, Matthies [\[34\]](#ref-cheng2006)), the robust-and-efficient stereo-feature-tracking method (Johnson et al. [\[67\]](#ref-johnson2008)), the attitude-and-position estimation account (Ali et al. [\[5\]](#ref-ali2006)), the FPGA-accelerated visual-odometry codesign for rover navigation (Lentaris et al. [\[73\]](#ref-lentaris2015)), the feature-point-selection scheme for stereo visual odometry (Motohashi and Kubota [\[89\]](#ref-motohashi2024)), the aided visual-odometry sparse-bundle-adjustment solution (Moore [\[87\]](#ref-moore2023); Moore [\[88\]](#ref-moore2023b)), the Perseverance path-estimation analyses (Andolfo, Petricca, Genova [\[6\]](#ref-andolfo2022); Andolfo [\[101\]](#ref-s2021); Andolfo [\[102\]](#ref-s2022)), and the precise stereo-vision pose estimation for Perseverance (Di et al. [\[41\]](#ref-di2022)) collectively establish how accurately the rover's traverse is known, which is the error budget the dependent variable inherits. The visual-odometry and visual-SLAM reviews (Yousif, Bab-Hadiashar, Hoseinnezhad [\[124\]](#ref-yousif2015); Aqel et al. [\[17\]](#ref-aqel2016); Fraundorfer and Scaramuzza [\[46\]](#ref-fraundorfer2012)) place these flight systems in the broader method literature and document the failure modes (feature-poor terrain, illumination) that bound the localization accuracy on which the distance measure rests. The path-planning review (Yang et al. [\[122\]](#ref-yang2023)) and the learning-based end-to-end lunar-rover path planning with safety constraints (Yu, Wang, Zhang [\[125\]](#ref-yu2021)) document the planning side of the autonomy stack whose autonomous operation the fraction measures, and the dynamic lateral-interaction model (Yeo [\[123\]](#ref-yeo2025)) and the scalable-bulldozing study ([\[22\]](#ref-c2025)) close the loop between the planning that the autonomy software performs and the wheel-soil mechanics that determine whether the planned motion is realized.

## 4.6 Data quality, validation against known values, and access and ethics

The final section treats the quality of the assembled measures, their validation against independently known values, and the access and ethics of the public archives. The case here is that the measurement design is auditable, that each principal construct can be validated against a value known from outside the construction, and that the use of these public archives raises no ethical or access barrier that would constrain the work. This rests on the documented accuracy of the localization pipeline, the documented behavioral statistics in the engineering reports, and the public, unrestricted, no-cost status of the three archives. A measure validated against an independently known value, and built from a source whose access is unrestricted, is a measure a third party can reproduce and check, and the localization and visual-odometry literature reports the accuracy figures against which the distance measure is validated. Validation bounds error but does not eliminate it, and the residual error is carried into the inference plan rather than assumed away. The remaining worry is that a clean-looking panel hides measurement error in the autonomous-fraction reconstruction, which the resolution indicator and the sensitivity analysis are designed to surface.

### 4.6.1 Validation of the dependent variable against known cumulative distances

The localized per-sol distances that the dependent variable is built from are validated against the cumulative traverse distances that the missions report independently. The mission overviews and campaign reports state cumulative distances at milestones (the kilometers-scale MER traverses in the mission overview, Crisp et al. [\[39\]](#ref-crisp2003); the Curiosity traverse through Gale crater, Vasavada et al. [\[110\]](#ref-vasavada2014), with the megaripple-crossing record to sol 710, Arvidson et al. [\[18\]](#ref-arvidson2017); the Perseverance first-campaign distances on the Jezero floor, Farley et al. [\[45\]](#ref-farley2023)). Summing the panel's localized per-sol distances over the corresponding span must reproduce these independently reported cumulative distances within the localization pipeline's error budget. This is a strong validation because the cumulative milestone distances are computed by the mission teams by a different aggregation than the panel's row-by-row sum, so agreement between the two is evidence that the row-level distances are correct and not merely internally consistent. The accuracy of the underlying localization is itself documented: visual odometry on MER converged on the large majority of attempts and measured position changes at the millimeter scale on steep slopes (Maimone, Cheng, Matthies [\[79\]](#ref-maimone2007)), and the Perseverance pose-estimation pipeline ties position to the orbital frame at high precision (Di et al. [\[41\]](#ref-di2022); Andolfo, Petricca, Genova [\[6\]](#ref-andolfo2022)). The confidence in the dependent variable is therefore rated high, conditional on this validation passing; the evidence that would lower it is a systematic discrepancy between the summed per-sol distances and the mission-reported cumulative distances on any one rover, which would signal a rover-specific localization bias that the panel would have to correct or flag.

### 4.6.2 Validation of the treatment measures

The autonomy-generation classification is validated against the TechPort TRL crosswalk of Section 4.3, which provides an external maturity timeline independent of the mission self-descriptions. The autonomous-drive fraction is validated, to the extent the record permits, against the aggregate behavioral statistics in the engineering reports: the MER tradeoff report's account of the daily mode choice and the roughly four-hour driving envelope (Biesiadecki, Leger, Maimone [\[20\]](#ref-biesiadecki2007)) bounds the plausible autonomous share for G1, and the Enhanced Navigation report's account of the raised autonomous share bounds it for G3 (Verma et al. [\[112\]](#ref-verma2025)). A reconstructed fraction that fell outside these documented envelopes would signal a reconstruction error. The chapter is candid that this validation is weaker than the distance validation, because the published behavioral statistics are aggregate rather than per-drive. That is why the autonomous-fraction construct carries moderate rather than high confidence and why the panel records a measurement-resolution indicator on each fraction value. The single most valuable pre-execution step, named here and in the prospectus, is the recovery of a per-drive command-mode record for each generation, which would convert the autonomous fraction from a reconstructed measure into a directly measured one and would raise its confidence to high.

### 4.6.3 Validation of the covariates

The hardware covariates are validated trivially, because wheel diameter, mass class, actuator class, and nominal drive energy are published constants of each rover in the mission descriptions (Crisp et al. [\[39\]](#ref-crisp2003); Grotzinger et al. [\[55\]](#ref-grotzinger2012); Maki et al. [\[80\]](#ref-maki2020)); there is no estimation error in a published constant, only the risk of transcription error, which a double-entry check against the source documents removes. The terrain covariates carry more error, because the terrain class and the physical-properties index are constructed by the candidate from orbital and early-sol characterizations (Golombek et al. [\[52\]](#ref-golombek2014); Arvidson et al. [\[18\]](#ref-arvidson2017)) rather than read from a published per-drive label, and the expansion plan flags the terrain-class crosswalk as a construction step rather than a citation. The validation here is internal-consistency checking against the terramechanics expectations: a terrain class labeled high-strength should not exhibit the realized-slip signatures that the terramechanics literature associates with weak soil (Ding et al. [\[42\]](#ref-ding2009); Ishigami et al. [\[63\]](#ref-ishigami2007)), and a mismatch would flag a misclassified class. Because realized slip is a mediator and cannot enter the model, this consistency check is used only for validating the terrain classification, never for adjusting the covariate inside the estimation, and the chapter records that distinction so the check cannot become a backdoor for a bad control.

### 4.6.4 Access and ethics

The access and ethics of the three archives are straightforward and are stated for completeness rather than because they pose a constraint. All three sources are public NASA archives. The Planetary Data System is the agency's permanent, openly accessible planetary-data archive; the NASA Technical Reports Server is the agency's public report index; NASA TechPort is the agency's public technology-portfolio record. None requires authentication, payment, or a data-use agreement for the research use described here, and the data are observational records of robotic spacecraft operations, so the work raises none of the human-subjects, privacy, or consent considerations that attend research on people. The principal ethical obligation in this setting is scientific rather than regulatory: the obligation to represent a public, taxpayer-funded data record accurately, to document the access path and construction so the panel is reproducible, and to label design-stage expectations as expectations rather than as findings. The dissertation discharges that obligation by recording every access path and construction procedure in the data-provenance appendix and by carrying the design-stage honesty rule into every measurement claim in this chapter. One forward-looking access caveat is worth noting: archive interfaces and product formats evolve, so the reproducibility of the panel depends on recording not only the access path but the retrieval date and product version, which the provenance appendix does.

### 4.6.5 Quality summary and how this chapter advances the argument

The measurement design makes two distinct contributions to the dissertation's argument, and it is worth stating them explicitly. First, it establishes that the question is measurable, by showing that the dependent variable, the treatment, and the covariates are each constructible from named public sources at the drive-sol unit of analysis. Second, it secures the causal mechanism at the measurement stage, by drawing the bad-controls line there so that realized slip and realized wheel-soil interaction, which are downstream of the autonomy software's choices both by the physics of terramechanics (Ding et al. [\[42\]](#ref-ding2009); Ishigami et al. [\[63\]](#ref-ishigami2007)) and by the econometric logic of post-treatment conditioning (Angrist and Pischke [\[7\]](#ref-angrist2009)), cannot enter the right-hand side. The residual measurement risk is concentrated in one construct, the autonomous-drive fraction, and it is bounded rather than hidden: the construct carries moderate confidence, the reconstruction procedure is stated in advance, a measurement-resolution indicator travels with each value, and the sensitivity of the autonomy estimate to measurement resolution is built into the analysis plan of Chapter 6. The chapter's overall confidence in the measurement design is rated high for the dependent variable and the hardware covariates, moderate for the terrain covariates and the generation indicator, and moderate for the autonomous-drive fraction, with the single highest-value path to raising the lowest grade, recovery of a per-drive command-mode record, named explicitly. This calibrated, source-anchored, mediator-excluding measurement design is the foundation on which the identification strategy of Chapter 5 is built, and it is constructed here so that the identification cannot later be undermined by an undocumented measurement choice.

The chapter has carried the prospectus data section forward and elaborated it from a three-source sketch into a complete operationalization: every dataset characterized by provenance, access, coverage, unit of analysis, and known biases; every variable given an operational definition, a source, and a scale in a single integrated table; the hardware and terrain covariates grounded in the terramechanics literature that also justifies the exclusion of the mediators; and the whole design checked for data quality, validated against independently known values, and cleared on access and ethics. What remains, and what the analysis plan and design chapters take up, is the estimation of the model these measures populate and the honest reporting of its uncertainty under the few-cluster structure that the panel's three-rover cross-section imposes.

## 4.7 Construct validity and external corroboration of the productivity measure

The sections above establish that the dependent variable is measurable, traceable, and internally consistent. They do not yet establish that it measures the right thing. That is the concern of construct validity: whether the operational measure, localized meters per sol or its inverse sols per meter, adequately captures the theoretical construct it is intended to represent, and whether the continuous autonomous-drive fraction adequately captures the construct of autonomy-driven performance. This section argues the case for each construct, names the threats to that case, proposes an external corroboration plan against an independently sourced measure, and commits to a measurement-reliability protocol for any coded inputs. No results are reported here; this is a plan that must be executed before the estimation of Chapter 5 can carry its intended meaning.

### 4.7.1 The construct defined precisely

The theoretical construct this dissertation seeks to measure is **traverse productivity**: the rate at which a rover converts a unit of mission time into useful surface coverage, under the operational and terrain conditions it actually faces. This construct has two components that must be distinguished. The first is the rate dimension: sols consumed per unit of distance, or distance covered per sol. The second is the mission-relevant component: the coverage must be coverage that advances the mission, which means it occurs on terrain the rover must actually traverse to reach science targets, not on arbitrarily chosen flat ground with no hazards and no scientific purpose.

The operational measures, meters per sol and sols per meter, capture the rate dimension directly and without ambiguity from the PDS localized path-length products (Di, Liu, and Yue [\[40\]](#ref-di2011); Carle, Furgale, and Barfoot [\[25\]](#ref-carle2010); Di et al. [\[41\]](#ref-di2022)). They do not capture the mission-relevance component. A rover instructed to drive in a circle on smooth ground would score well on meters per sol without advancing the mission at all. The construct this dissertation targets is therefore narrower than science return and broader than raw odometry: it is traverse productivity along operationally real routes, on the terrain each rover actually drove, toward targets the mission actually selected.
Three things the construct is intended to capture: the rate at which the rover executes commanded traverses across the terrain classes it encountered; the degree to which onboard autonomy, by handling hazards in real time, raises that rate relative to blind commanded driving on the same terrain; and the interaction between autonomy capability and terrain difficulty, which is the mechanism prediction that separates the autonomy explanation from a simple mechanical-size explanation.

Three things the construct explicitly omits and must not be read as capturing: the scientific value of the targets reached, which depends on target selection rather than traverse efficiency; the planning effort the ground team expends per meter driven, a real operational cost that the simple distance-and-sols measure ignores (Farley et al. [\[45\]](#ref-farley2023); Gao and Chien [\[50\]](#ref-gao2021)); and absolute distance accumulated across the mission lifetime, which conflates mission duration with per-sol rate. The dissertation's chosen operationalization, the per-sol rate and its inverse, captures what is directly and consistently available in the public archive, remains uncontaminated by the analyst's model of target-selection or ground-team behavior, and flags the richer construct as future work, consistent with the delimitation stated in Section 1.7.

### 4.7.2 Threats to construct validity of the productivity measure

Four threats to the construct validity of meters per sol as a measure of traverse productivity are named here. Three are general; one is specific to the comparison across generations.

**Threat 1: terrain confounding of raw distance.** Raw distance per sol is not a pure measure of rover capability. It is also a measure of how cooperative the terrain was. A rover that drives five hundred meters on flat bedrock is no more productive, in any meaningful sense, than one that drives two hundred meters across a megaripple field, yet the raw-distance measure ranks the first higher. The mitigation is the terrain-class fixed effect \(\delta_{c}\) and the continuous terrain covariates (slope and physical-properties index), included in the estimating equation precisely to remove systematic between-terrain differences from the productivity comparison (Golombek et al. [\[52\]](#ref-golombek2014); Arvidson et al. [\[18\]](#ref-arvidson2017)). After conditioning on terrain class and covariates, the residual variation in meters per sol approaches a pure capability measure. The residual threat is that terrain is measured with error and that the crosswalk from orbital basemaps to drive-level class is imperfect; this is noted as a moderate-confidence construction in Section 4.6.3 and carried into the sensitivity analysis of Chapter 5.

**Threat 2: planning-cycle and uplink-cadence drivers of sols per meter.** The sols-per-meter inverse measure can rise (deteriorate) not because the rover is less capable but because the ground team commands shorter drives on a given sol, either to stay conservative around a newly characterized hazard or because the planning cycle ran short. A sol on which the rover could drive five hundred meters but is commanded to drive fifty meters scores worse on sols per meter than its capability warrants. This confound operates at the sol level and is partially, not fully, addressed by the within-rover autonomous-fraction contrast: the within-rover comparison includes sols of varying commanded distance and therefore partly averages over planning-cadence variation. The partial mitigation is to use the commanded-drive parameters as a priori controls (they are pre-treatment quantities and are eligible covariates per the bad-controls classification of Section 4.4), so that variation in commanded distance within a sol is conditioned out. The residual threat is that the ground team's conservatism is itself shaped by terrain assessment, making commanded distance partially endogenous to terrain; terrain-class conditioning is therefore a necessary complement to the drive-parameter controls rather than an alternative to them.

**Threat 3: hardware-autonomy collinearity as a construct threat.** The construct targeted is the contribution of the autonomy channel to traverse productivity, holding hardware fixed. The threat is that across rovers the autonomy generation is perfectly collinear with the hardware generation: every new rover upgraded both at once, so a between-rover comparison of productivity cannot cleanly attribute the difference to autonomy rather than to the larger wheel or the stronger actuator. This threat is not unique to measurement; it is the central identification problem of the dissertation and is treated at length in Sections 1.3 and 5.4 (cross-reference). It is named here as a construct-validity threat because it bears on whether the between-rover productivity comparison, even when accurately measured, measures what it claims to measure. The mitigation at the construct level is the continuous autonomous-drive fraction, which varies within rover and therefore, in the within-rover specification, measures the autonomy contribution while holding the hardware platform constant at the level of the individual machine.

**Threat 4: localized path length versus effective progress.** The dependent variable is built from localized path length, the actual path the rover drove, not net displacement. On a route that doubles back around a hazard, path length exceeds displacement, and the rover is credited with the extra distance the detour required. Section 4.4 argues that this is the correct measure, because it rewards the rover for the ground actually covered rather than for straight-line net progress, and because a rover that detours autonomously around a hazard without stopping is performing the very capability the autonomy channel claims to deliver. The construct-validity risk is that a rover taking longer, less direct routes scores higher than one taking shorter direct routes on the same terrain, even when both traverse the same scientific ground. The mitigation is terrain conditioning, which removes the systematic route-length differences across terrain classes, and the within-rover comparison, which holds route-selection behavior approximately constant at the level of the individual machine.

### 4.7.3 External corroboration plan

Construct validity for a survey or index measure is conventionally supported by convergent validity, the demonstration that the measure correlates with a second, independently sourced measure of the same underlying construct, and by discriminant validity, the demonstration that it does not correlate too highly with measures of a different construct (Angrist and Pischke [\[7\]](#ref-angrist2009)). For the productivity measure used here, convergent validity requires an independent, separately sourced measure that reflects traverse efficiency rather than raw distance or raw sol count.

The proposed external corroboration measure is **science-target acquisition rate per sol**: the number of distinct science targets reached and characterized on a rolling basis, divided by the sols required to reach them, drawn from the publicly available science-campaign records of each mission. For the Mars Exploration Rovers, the science-campaign records document instrument placements and the rock and soil targets engaged on each sol (Crisp et al. [\[39\]](#ref-crisp2003); Squyres et al. [\[106\]](#ref-squyres2006)). For Perseverance, the cache-target and sample-acquisition record provides an analogous series: each sample tube sealed represents a target reached and characterized, and the sol count to each sealing event is a sol-level productivity observation that does not depend on the localized-distance measure (Farley et al. [\[45\]](#ref-farley2023); Verma et al. [\[111\]](#ref-verma2023)). This science-target-acquisition rate is independently sourced, because it is drawn from the science operations record rather than the traverse-product archive; it is separately constructed, because it counts events rather than distances; and it captures a mission-relevant dimension of productivity that the simple distance-and-sols measure explicitly omits.

The convergent expectation is that per-sol drive-distance productivity and per-sol science-target-acquisition rate should be positively correlated within rover across comparable terrain classes: on sols when the rover covers more ground per sol, it should also reach more targets per sol, because traverse efficiency and target acquisition are driven by the same underlying capability. A Pearson correlation between the two series, computed within rover after removing terrain-class means, is the primary convergent check. A correlation substantially above zero and consistent in sign across the three full panel members would support the claim that the distance-and-sols measure tracks the same underlying construct as the science-operations record.

The discriminant expectation is that the distance-based productivity measure should correlate only weakly with the sol-level science-instrument contact time (the time a given instrument spends engaged with a target), which measures science intensity rather than traverse efficiency. A high correlation between distance productivity and instrument-contact time would suggest the distance measure is picking up science campaigns rather than traverse capability, a form of construct contamination. A low correlation would support discriminant validity.

A divergence between the distance-based measure and the science-target-acquisition rate would be informative in a specific direction. If the two measures diverge, the most likely explanation is that the ground team substituted instrument time for drive time on productive sols, so that high-productivity sols are allocated partly to science activities and therefore show fewer targets per sol despite more ground covered. This divergence would confirm the planning-effort omission flagged in the construct definition of Section 4.7.1, and would mean the distance-and-sols measure understates the operational benefit of autonomy by omitting the planning-cycle freeing effect. The dissertation states this possibility explicitly in Section 1.6.2 and treats it as a conservative bias: a confirmed autonomy effect on the distance measure would, if this divergence holds, understate the full value of the autonomy channel.

The corroboration plan is not a substitute for the main estimation. It is a check that, if it passes, strengthens the claim that the distance-based measure tracks a real underlying construct rather than an artifact of mission duration or terrain assignment. If it fails, the dissertation must qualify the productivity construct more narrowly and restrict the scope of its conclusions to the distance-and-sols measure specifically.

### 4.7.4 Measurement-reliability protocol for coded inputs

The construction of the autonomous-drive fraction, described in Section 4.2.3, requires classifying individual drive segments as either blind commanded motion or autonomous navigation. This classification is a coding decision made by the researcher on the basis of the command and telemetry mode record. Any coding decision carries the risk of inter-coder inconsistency: two researchers applying the same procedure to the same command record may classify the same segment differently, introducing measurement error in the autonomous-fraction series. This risk is non-trivial because the command records for some generations, particularly for ambiguous segments where the commanding mode is not explicitly flagged as autonomous or blind, demand judgment calls about which classification rule applies.

The measurement-reliability protocol for this coded input runs as follows. The coding scheme for segment classification must be written out in advance as a decision rulebook, specifying the exact criteria by which a segment is classified as autonomous (onboard hazard detection and path planning active, no ground-specified path for the segment) or as blind commanded (ground-specified path, onboard system in position-hold or dead-reckoning mode only). The rulebook is a pre-registered document, completed before any segment is coded, and filed as part of the pre-registration record of Section 5.8.

A second coder, working independently from the same command and telemetry records and the same rulebook, re-derives the autonomous-fraction classification for a randomly drawn subsample of drive segments. The subsample size must suffice to estimate a Cohen's kappa statistic with a confidence interval narrow enough to distinguish acceptable from unacceptable agreement; an illustrative target, stated here as a design expectation rather than an executed result, is a subsample of at least fifty segments per rover per generation, yielding a kappa interval with a half-width below 0.10 at the ninety-five-percent level. The pre-registered acceptability threshold is a kappa of 0.80 or above, which corresponds to strong agreement in the standard classification and marks the boundary the measurement-reliability literature commonly treats as the line between moderate and high reliability.

If the kappa falls below the threshold for any rover-generation cell, the coding rulebook is revised, both coders re-apply the revised rules to the full disagreement set, and the revised kappa is reported alongside the original. The final autonomous-fraction series is constructed from the primary coder's classifications on all segments, with the second-coder disagreement rate reported as a quality indicator in the provenance appendix. The measurement-resolution indicator introduced in Section 4.2.3 is updated to reflect the inter-coder agreement grade: segments classified consistently by both coders receive a high-resolution indicator; segments where the two coders disagreed receive a low-resolution indicator, and the sensitivity of the within-rover autonomy coefficient to dropping low-resolution segments is reported as one dimension of the robustness battery of Section 5.6.

This protocol does not eliminate measurement error in the autonomous-fraction series; it bounds and discloses it. A fraction series with strong inter-coder agreement can still carry systematic error if both coders apply the same misinterpretation of an ambiguous telemetry field. The mitigation against systematic error is the validation against the behavioral-statistics envelopes of the engineering reports described in Section 4.6.2: a fraction series that passes both the inter-coder agreement check and the behavioral-statistics validation is a series whose reliability is supported at two independent levels, and whose residual measurement error is the smaller classical component that attenuates the autonomy coefficient toward zero and therefore works against H1 rather than inflating it.

### 4.7.5 How this section advances the argument

This section closes the pre-defense gap between measurement and construct. Sections 4.1 through 4.6 established that the productive measures are computable, traceable, and internally validated. This section argues that they measure what the dissertation claims they measure, names the ways in which that claim could fail, and proposes the verification steps that would support or undermine it. The construct definition of Section 4.7.1 binds the rest of the dissertation: every interpretation of a coefficient in Chapters 5 and 6 should be read as an interpretation about traverse productivity in the specific sense defined there, not about science return, mission success, or the broader value of autonomous systems. The external corroboration plan of Section 4.7.3 is an empirical commitment that can fail, and naming it in advance is the discipline that separates a construct argument from an asserted one. The measurement-reliability protocol of Section 4.7.4 is a procedural commitment, verifiable before estimation begins, whose outcome travels into the sensitivity analysis as a documented quality indicator rather than an undisclosed assumption. Together, the four subsections above convert the productivity measure from a convenient archival quantity into a construct-grounded dependent variable whose validity is argued rather than assumed.



# Chapter 5: Research Design and Identification
## 5.1 The chapter thesis and the problem it answers

The two-way fixed-effects panel with a within-rover autonomous-fraction contrast is chosen because it is the only specification available in this setting that holds the mechanical platform fixed at the level of the individual machine while letting the autonomy contribution vary. The credibility of the whole dissertation stands or falls on two disciplines applied to that specification: a strict bad-controls rule that keeps every quantity produced by the drive on the outcome side, and an inference procedure honest about the small number of independent rover clusters. That is the claim this chapter defends, and the sections below develop it in turn.

The problem this chapter addresses can be stated precisely. The question of what drives Mars surface mobility productivity, set out in Chapter 1 and motivated by the two competing narratives in Chapters 2 and 3, was translated in Chapter 4 into a set of measurable constructs: a dependent variable (localized meters per sol and its inverse, sols per meter), a categorical and a continuous treatment (autonomy-software generation G1 through G3, and the autonomous-drive fraction), a block of hardware covariates, and a terrain block that serves both as covariate and as fixed-effect dimension. What remains is to supply an estimator and an identification argument that recover the causal contribution of the autonomy channel net of the mechanical channel, with the threats to that recovery named and mitigated in advance. The constructs alone do not identify anything. Autonomy generation and hardware generation move together across the fleet by construction, because every new rover upgraded both at once, so a naive regression of productivity on autonomy generation cannot tell whether the coefficient reflects software, steel, or the correlation between them. Left unaddressed, this would reproduce in numerical form exactly the combined narrative the dissertation set out to improve upon: it would credit later rovers with higher productivity without separating the two channels, and the resulting coefficient would be uninterpretable as a causal effect of either one. This chapter resolves that difficulty by specifying the estimator, arguing the identification assumptions formally, and laying out the apparatus of threat mitigation, robustness, power analysis, pre-registration, and computation that turns a set of constructs into a falsifiable design.

The identification strategy carries forward the commitments the earlier chapters established. The productivity rise is genuine, and the two channels have been described together and never separated (Crisp et al. [\[39\]](#ref-crisp2003); Grotzinger et al. [\[55\]](#ref-grotzinger2012); Vasavada et al. [\[110\]](#ref-vasavada2014); Farley et al. [\[45\]](#ref-farley2023); Gao and Chien [\[50\]](#ref-gao2021); Verma et al. [\[112\]](#ref-verma2025)). The separation matters because sols are the binding scarcity for the sampling and caching activity that defines the Mars Sample Return campaign, so a meter saved is a sol freed (Farley et al. [\[45\]](#ref-farley2023); Golombek et al. [\[52\]](#ref-golombek2014); Genova et al. [\[51\]](#ref-genova2013)). The design reaches the causal mechanism directly: a within-rover autonomous-fraction contrast holds hardware fixed and isolates the software channel, while the bad-controls rule keeps realized slip on the outcome side (Angrist and Pischke [\[7\]](#ref-angrist2009); Angrist and Pischke [\[8\]](#ref-angrist2014); Maimone, Cheng, and Matthies [\[79\]](#ref-maimone2007); Biesiadecki and Maimone [\[20\]](#ref-biesiadecki2007); Helmick et al. [\[59\]](#ref-helmick2004)). It improves on the alternatives because the within-rover contrast dominates both the combined-narrative description and the naive between-rover regression that cannot separate collinear channels (Angrist and Pischke [\[10\]](#ref-angristpischke2010); de Chaisemartin and D'Haultfoeuille [\[29\]](#ref-chaisemartin2020); Goodman-Bacon [\[54\]](#ref-goodmanbacon2021); Callaway and Sant'Anna [\[23\]](#ref-callawaysantanna2021)). The risk that remains is bounded and visible: wild-cluster bootstrap inference and an explicit threat-to-validity treatment make the residual identification risk transparent (de Chaisemartin and D'Haultfoeuille [\[28\]](#ref-chaisemartin2019); Goodman-Bacon [\[53\]](#ref-goodmanbacon2018); Imbens and Angrist [\[60\]](#ref-imbensangrist1994); Leamer [\[72\]](#ref-leamer2010)). The remainder of the chapter develops each of these in turn.

A note on scope is owed at the outset and binds everything below. This is a design-stage dissertation. No estimation has been executed, and no coefficient reported here is an estimate produced from the assembled panel. Where a number appears, it is an illustrative expectation or a calculation made to size the design (for example, a minimum-detectable-effect figure computed from assumed variances), and it is labeled as such. The contribution at this stage is the design itself, argued to the standard of a pre-registration, not a result.

## 5.2 The estimator and why it is chosen

### 5.2.1 The baseline specification

The estimator is a panel linear regression with two-way fixed effects. The specification is reproduced here verbatim from the notation fixed in the prospectus and the design specification, because consistency across chapters depends on the symbols carrying identical meaning in Chapter 5, where they are defended, and in Chapter 6, where the analysis plan operates on them. For drive-sol *i* on rover *r* in terrain class *c*:

\[
\text{Productivity}_{irc} = \beta_1 \, \text{AutonomyGen}_{r}
                 + \beta_2 \, \text{Hardware}_{r}
                 + \gamma \, \text{Terrain}_{ic}
                 + \alpha_{r} \quad \text{(rover fixed effect)}
                 + \delta_{c} \quad \text{(terrain-class fixed effect)}
                 + \epsilon_{irc} \qquad\qquad (1)
\]

Each term carries a precise role. \(\text{Productivity}_{irc}\) is localized meters per sol in the primary specification and sols per meter in the parallel specification, computed from the Planetary Data System traverse product as localized path length rather than straight-line displacement, so the measure reflects the distance actually driven rather than net progress toward a goal. \(\text{AutonomyGen}_{r}\) is, in the categorical form, the generation indicator (G1 for the Mars Exploration Rover-class AutoNav with visual odometry, G2 for the Mars Science Laboratory inherited and extended stack, G3 for the Mars 2020 Enhanced Navigation), and in the continuous secondary form it is the autonomous-drive fraction, the share of a drive's localized distance executed under onboard autonomous navigation rather than blind commanded motion. \(\text{Hardware}_{r}\) is the block of mechanical covariates (wheel diameter, mass class, actuator class, nominal available drive energy per sol). \(\text{Terrain}_{ic}\) is the continuous terrain block (slope and a physical-properties index) that varies within terrain class. The term \(\alpha_{r}\) is the rover fixed effect, absorbing all fixed rover-specific characteristics; \(\delta_{c}\) is the terrain-class fixed effect, absorbing systematic differences between terrain classes; and \(\epsilon_{irc}\) is the idiosyncratic disturbance.

### 5.2.2 Why a two-way fixed-effects panel rather than a cross-section or a single time series

The two-way fixed-effects panel dominates the available alternatives in this setting, and the case for it rests on the structure of the data and the structure of the confounding. The governing principle is the design-based one: the estimator should be the one that makes the most defensible counterfactual comparison given the variation actually present, not the most flexible functional form available in the abstract (Angrist and Pischke [\[7\]](#ref-angrist2009); Angrist and Pischke [\[10\]](#ref-angristpischke2010)). That principle is borne out by the body of applied work in which the choice of estimator turns on which comparison it implicitly makes, and in which the credibility of an estimate is judged by the transparency of that comparison rather than by the sophistication of the model (Angrist and Pischke [\[8\]](#ref-angrist2014); Leamer [\[72\]](#ref-leamer2010)).

Consider the alternatives the panel improves upon. A pure cross-section of rovers, regressing productivity on autonomy generation across the four (effectively three full) flight rovers, would have a sample of three or four observations at the rover level and could identify nothing, because autonomy generation, hardware generation, and landing-site terrain would be perfectly or near-perfectly collinear at that level of aggregation. A single rover's time series, regressing its drive-sol productivity on time, would conflate autonomy with the operational learning of the ground team and with the changing terrain along the traverse, and would have no leverage on the cross-generation question at all. The panel is the structure that lets the two fixed-effect dimensions do their work: \(\alpha_{r}\) removes everything fixed about a rover (its landing site's gross character, its instrument suite, its team's standing procedures), so that the autonomy effect, where it is identified within rover, is identified from variation that the rover's own constants cannot explain; and \(\delta_{c}\) removes the systematic productivity differences between terrain classes, so that the comparison is between drives on comparable ground.

One qualification is essential and is protected rather than buried. The two-way fixed-effects panel dominates the alternatives only for the within-rover identification of the autonomous-fraction effect and the terrain-conditioned between-rover identification of the generation effect; it does not dissolve the fundamental collinearity of autonomy and hardware at the between-rover level, which no estimator applied to this fleet can fully dissolve because the two were upgraded together every mission. The objection the design must answer is therefore not that some other estimator would do better, but that the fixed-effects estimator, like any estimator here, leans on identifying assumptions that must be argued rather than assumed. Those assumptions are the subject of Section 5.4.

### 5.2.3 Why not an instrumental-variables or difference-in-differences design

Two estimators that the methodological anchors are famous for are deliberately not adopted as the primary design, and saying why sharpens the choice that is made. An instrumental-variables design would require an instrument that shifts autonomy generation without affecting productivity except through autonomy, satisfying the exclusion restriction and the monotonicity condition that define a local average treatment effect (Imbens and Angrist [\[60\]](#ref-imbensangrist1994); Angrist, Imbens, and Rubin [\[9\]](#ref-angristimbensrubin1996)). No such instrument exists in this setting. The plausible candidates (mission funding date, processor availability, the calendar of flight-software releases) all plausibly affect productivity through channels other than autonomy generation, for example through contemporaneous hardware decisions or through the maturity of ground operations, so the exclusion restriction would fail. The design therefore does not claim an instrument, and it does not manufacture one; where the corpus and the setting lack the variation needed for a credible instrument, the honest course is to say so and to identify the effect by the within-unit comparison instead. This is itself an application of the credibility discipline: a weak or invalid instrument is worse than no instrument, because it lends a spurious causal gloss to a correlation (Angrist and Pischke [\[10\]](#ref-angristpischke2010)).

A difference-in-differences design with staggered adoption would treat each rover's transition to a new autonomy generation as a treatment timed differently across units, and the recent econometric literature on exactly this case informs the design (Goodman-Bacon [\[53\]](#ref-goodmanbacon2018); Goodman-Bacon [\[54\]](#ref-goodmanbacon2021); de Chaisemartin and D'Haultfoeuille [\[28\]](#ref-chaisemartin2019); de Chaisemartin and D'Haultfoeuille [\[29\]](#ref-chaisemartin2020); Callaway and Sant'Anna [\[23\]](#ref-callawaysantanna2021)). That literature is drawn on here not because the dissertation runs a staggered difference-in-differences, but because it supplies two lessons that the fixed-effects design must absorb. The first is that the two-way fixed-effects estimator, when treatment effects are heterogeneous across units and time, can produce a weighted average of treatment effects with some negative weights, so that the headline coefficient is not a clean average of the underlying effects (de Chaisemartin and D'Haultfoeuille [\[29\]](#ref-chaisemartin2020); Goodman-Bacon [\[54\]](#ref-goodmanbacon2021)). The second is the constructive response: estimate the heterogeneous effects directly, by interacting the treatment with terrain class and by examining the effect by generation rather than forcing a single homogeneous coefficient (Callaway and Sant'Anna [\[23\]](#ref-callawaysantanna2021)). The dissertation adopts the lessons (heterogeneity is estimated, not assumed away) without adopting the staggered-adoption framing, because the autonomous-drive fraction is a continuous within-rover treatment rather than a discrete adoption event, and the within-rover continuous contrast is the cleaner identification. This literature is named for what it inherits methodologically, not for the estimator it would imply, and the distinction is drawn explicitly so the reader does not mistake the citation for a claim that a difference-in-differences is being run.

## 5.3 The specifications written out

### 5.3.1 The two outcome operationalizations in parallel

Mobility productivity is estimated in two parallel forms, and both are run rather than one being chosen, because each answers a slightly different question and the pair guards against an artifact of functional form. The meters-per-sol specification puts the rate of distance production on the left-hand side and asks what raises that rate. The sols-per-meter specification puts the inverse on the left-hand side and asks what lowers the sol cost of a meter, which is the quantity a mission planner actually budgets in the daily planning cycle. The two are not mechanically redundant: a regression linear in meters per sol is not linear in sols per meter, so a covariate can appear significant in one and attenuated in the other, and the divergence between them is informative about where the relationship is nonlinear. Both are needed because the planner's decision is framed in sols per meter while the engineering literature reports meters per sol, and a design meant to be decision-relevant must report the quantity the decision-maker uses while remaining comparable to the literature it builds on (Gao and Chien [\[50\]](#ref-gao2021); Farley et al. [\[45\]](#ref-farley2023)). Where the two specifications agree on sign and significance, confidence in the result is higher; where they diverge, the divergence is reported as a finding about nonlinearity rather than reconciled by privileging one form.

### 5.3.2 The within-rover specification (primary)

The primary test of H1 is the within-rover autonomous-fraction specification. It retains the rover fixed effect \(\alpha_{r}\) and uses the continuous autonomous-drive fraction as the treatment:

\[
\text{Productivity}_{irc} = \beta_{\text{af}} \, \text{AutonomousFraction}_{irc}
                 + \gamma \, \text{Terrain}_{ic}
                 + \alpha_{r} \quad \text{(rover fixed effect)}
                 + \delta_{c} \quad \text{(terrain-class fixed effect)}
                 + \epsilon_{irc} \qquad\qquad (2)
\]

Here the hardware block drops out of the right-hand side, not because hardware is irrelevant but because it is constant within a rover and is therefore absorbed entirely by \(\alpha_{r}\). This is the specification's central virtue. Within a single machine, wheel diameter, mass, actuators, and nominal energy do not vary from one drive-sol to the next; the only thing that varies between a segment driven autonomously and a segment driven under blind command is how much of the driving the software is doing. The coefficient \(\beta_{\text{af}}\) is therefore identified from variation that holds the entire mechanical platform fixed by construction, which is the closest this setting can come to the design-based ideal of comparing like with like (Angrist and Pischke [\[8\]](#ref-angrist2014)). The terrain block and the terrain-class fixed effect remain on the right-hand side because the autonomous fraction is not assigned at random within rover; the ground team chooses to drive autonomously when it judges the terrain suitable and the schedule pressing, so the autonomous fraction is correlated with terrain, and terrain affects productivity directly. Conditioning on terrain is what makes the residual variation in autonomous fraction approach as-good-as-random with respect to productivity, and this conditioning is argued formally in Section 5.4.

### 5.3.3 The between-rover specification (secondary)
The secondary test relaxes the rover fixed effect to a terrain-class fixed effect plus the hardware block, so that the categorical generation contrast becomes identifiable from between-rover variation:

\[
\text{Productivity}_{irc} = \beta_1 \, \text{AutonomyGen}_{r}
                 + \beta_2 \, \text{Hardware}_{r}
                 + \gamma \, \text{Terrain}_{ic}
                 + \delta_{c} \quad \text{(terrain-class fixed effect)}
                 + \epsilon_{irc} \qquad\qquad (3)
\]

This specification cannot retain \(\alpha_{r}\). A rover fixed effect that fully absorbs rover identity would also absorb autonomy generation, which is a rover-level attribute, and leave \(\beta_1\) unidentified. The price of identifying the generation effect between rovers is that the hardware block must carry the burden of separating the autonomy channel from the mechanical channel, and that burden is heavy precisely because the two are collinear. The between-rover specification is reported as a complement to the within-rover specification, never as a substitute for it. The two together bound the effect of interest. The within-rover estimate is conservative and clean because hardware is held fixed at the level of the machine; the between-rover estimate is broader in scope because it speaks to whole-generation differences, but it leans on the hardware controls and is correspondingly more fragile. Presenting both, and stating which is the load-bearing identification, is the design's answer to the collinearity problem rather than a concession to it.

### 5.3.4 The nested decomposition

To make the relative contribution of the two channels legible, the between-rover specification is estimated in three nested forms, each adding a block:

1. Terrain fixed effects only.
2. Terrain fixed effects plus the hardware block.
3. Terrain fixed effects plus the hardware block plus the autonomy block.

The quantity of interest is the incremental explanatory share contributed by the autonomy block in moving from form (2) to form (3), compared against the incremental share contributed by the hardware block in moving from form (1) to form (2). Under H1 the autonomy block's incremental share is the larger of the two, and the hardware coefficients attenuate when autonomy enters, indicating that part of what the hardware block absorbed in form (2) was the correlated autonomy upgrade. Under H0 the autonomy block adds little once hardware is in, and the hardware coefficients are stable across forms (2) and (3). The nested decomposition is a description device, not by itself an identification. Its interpretation is governed by the identification argument of Section 5.4 and by the bad-controls rule, because the incremental-share comparison is meaningful only if neither block contains a post-treatment mediator.

## 5.4 Identification assumptions argued formally

### 5.4.1 The identifying variation and the counterfactual it implies

Identification of H1 against H0 rests on naming the counterfactual and the comparison that recovers it, which is the first discipline the design-based apparatus imposes (Angrist and Pischke [\[7\]](#ref-angrist2009); Angrist and Pischke [\[10\]](#ref-angristpischke2010)). The counterfactual of interest is the productivity a given rover would have achieved on a given drive-sol had it executed a larger or smaller share of that drive autonomously, terrain held fixed. The comparison that approximates this counterfactual is the within-rover, within-terrain-class variation in the autonomous-drive fraction. The identifying assumption, stated formally, is conditional mean independence: after conditioning on the rover fixed effect, the terrain-class fixed effect, and the continuous terrain block, the autonomous-drive fraction is mean-independent of the disturbance \(\epsilon_{irc}\). In symbols, the assumption is that the expectation of \(\epsilon_{irc}\) given \(\text{AutonomousFraction}_{irc}\), \(\text{Terrain}_{ic}\), \(\alpha_{r}\), and \(\delta_{c}\) does not depend on \(\text{AutonomousFraction}_{irc}\). If this holds, \(\beta_{\text{af}}\) is the causal effect of the autonomous fraction on productivity for the population of drive-sols represented in the panel.

This assumption is more defensible in the within-rover specification than in any cross-rover alternative, because the largest confounders of the autonomy-productivity relationship are rover-level and terrain-level: the mechanical platform, the instrument suite, the landing site's gross terrain character, the team's standing procedures. The rover fixed effect removes every rover-level confounder by construction, and the terrain blocks remove the terrain-level confounders down to within-class residual variation. What makes this work is the fixed-effects logic that a confounder constant within the unit of comparison cannot bias a within-unit estimate (Angrist and Pischke [\[7\]](#ref-angrist2009)). The long applied record of within-unit designs recovering credible effects where between-unit designs cannot bears it out, because the within-unit comparison differences out the time-invariant heterogeneity that drives the between-unit confounding (Angrist and Pischke [\[8\]](#ref-angrist2014); Angrist and Pischke [\[10\]](#ref-angristpischke2010)).

One qualification is protected throughout: conditional mean independence is an assumption, not a fact, and it can fail. It fails if, within a rover and within a terrain class, the ground team chooses higher autonomous fractions on drive-sols that would have been more productive anyway for reasons not captured by the terrain block. The design's answer to this objection is twofold. First, the terrain block is constructed to be rich enough that the residual selection of autonomous fraction is plausibly minor: slope and a physical-properties index, conditioned on terrain class, capture the main drivers of the team's autonomy decision. Second, the direction of any residual selection is argued in Section 5.4.3 to bias the estimate toward the null rather than away from it, which makes a positive finding conservative. Neither point makes the assumption true; both make a positive result harder to dismiss as selection.

### 5.4.2 The bad-controls rule, stated and applied

The second identification discipline is the bad-controls rule: do not condition on variables that are themselves outcomes of the treatment, because doing so absorbs part of the effect of interest and biases the estimate, typically toward the null (Angrist and Pischke [\[7\]](#ref-angrist2009)). The application here is sharp and consequential. Realized slip, realized wheel-soil interaction, and realized drive duration are all produced by the drive itself, and they are produced in part by how the autonomy software chose to drive. A rover driving autonomously may select a path that reduces slip, or may take longer per meter because it pauses to assess hazards; either way, slip and duration are post-treatment quantities that lie on the causal path between the autonomous fraction and productivity. Conditioning on them would block part of the very mechanism the design seeks to measure.

The design therefore draws a clean and pre-committed line. A priori terrain class, a priori terrain covariates (slope, physical-properties index from orbital basemaps), and commanded-drive parameters are eligible controls, because they are determined before or independently of the drive's execution. Anything realized during or after the drive (slip, wheel sinkage, actual duration, energy actually consumed) is a mediator and stays on the outcome side, never entering the right-hand side. This line is not a refinement that can be relaxed for convenience; it is constitutive of the identification, because the estimand is the total effect of the autonomy channel on productivity, which includes its effect through slip avoidance and hazard navigation, and conditioning on the mediators would redefine the estimand as a direct effect net of those channels (Angrist and Pischke [\[7\]](#ref-angrist2009); Angrist and Pischke [\[8\]](#ref-angrist2014)). The mechanism literature supports treating slip as a consequence of the driving decision rather than a pre-determined feature of the world: visual odometry exists precisely to let the rover measure and respond to slip during the drive, which makes slip an output of the autonomy system, not an input to it (Maimone, Cheng, and Matthies [\[79\]](#ref-maimone2007); Helmick et al. [\[59\]](#ref-helmick2004)).

### 5.4.3 The direction of residual selection bias

A causal claim that names a mechanism is stronger than one that asserts a bare correlation, so the design states the mechanism by which residual selection would bias \(\beta_{\text{af}}\) and signs that bias. The driver is the ground team's autonomy decision. The mechanism is that the team tends to authorize higher autonomous fractions on harder, more hazard-dense terrain where blind commanded driving is slow and risky, and tends to drive blind on benign, well-imaged flats where commanded motion is already near-optimal. The observable effect is a negative correlation between autonomous fraction and the easy-terrain productivity that blind driving achieves. The operational consequence is that, absent perfect terrain conditioning, the autonomous fraction would appear on the right-hand side partly as a proxy for difficult terrain, which is low-productivity terrain. The implication for the design is that any imperfectly conditioned residual selection pushes \(\beta_{\text{af}}\) downward, toward and possibly below zero, not upward. Conditioning on terrain class and the continuous terrain block is intended to remove this; to the extent it does not fully succeed, the bias works against H1. A positive, significant \(\beta_{\text{af}}\) is therefore conservative with respect to this threat, because the threat would have suppressed it. This mechanism-named, direction-signed reasoning converts a hand-waved worry into a manageable, argued residual risk. It raises confidence in a positive result while conceding that a null result could reflect either a true absence of effect or an over-correction by the terrain controls.

### 5.4.4 The collinearity problem and the two-pronged resolution

The central internal-validity threat, named repeatedly because it is the one the whole design is built around, is the collinearity of autonomy generation and hardware generation. Each new rover upgraded both at once: Curiosity was both a larger machine than the Mars Exploration Rovers and ran an extended autonomy stack; Perseverance was both a still larger machine and introduced Enhanced Navigation (Grotzinger et al. [\[55\]](#ref-grotzinger2012); Maki et al. [\[80\]](#ref-maki2020); Verma et al. [\[112\]](#ref-verma2025)). At the rover level the two are nearly perfectly collinear, and no regression on three or four rovers can separate them.

The resolution is two-pronged, which is the reason the design has two specifications rather than one. The first prong is the within-rover autonomous-fraction contrast, which sidesteps the collinearity entirely by holding hardware constant at the level of the individual machine; within a rover there is no hardware variation to be collinear with, so \(\beta_{\text{af}}\) is identified from autonomy variation alone. The second prong is the between-rover generation contrast with the hardware block explicitly included, which does not sidestep the collinearity but confronts it, accepting that the hardware controls carry a heavy burden and that the resulting generation coefficient is more fragile. The two prongs together are more informative than either alone. The within-rover prong establishes that the autonomy channel has a real, hardware-free effect; the between-rover prong, read in light of the within-rover result, indicates whether that effect scales up to whole-generation differences or whether the between-rover difference is dominated by hardware. The logic that justifies pairing them is that a clean lower-leverage estimate and a broader higher-risk estimate jointly bracket the parameter more tightly than a single estimate of either kind (Angrist and Pischke [\[8\]](#ref-angrist2014); de Chaisemartin and D'Haultfoeuille [\[29\]](#ref-chaisemartin2020)). One qualification is that the between-rover prong remains unidentified in the strict sense and is interpreted as suggestive, with its standard errors computed by the few-cluster procedure of Section 5.5 and its caveat stated wherever it is reported.

## 5.5 Threats to validity and their mitigations

This section enumerates the threats to validity across the four standard categories and pairs each with its mitigation, so the residual risk is transparent. The organizing claim is that each threat is either mitigated by a feature of the design or, where it cannot be fully mitigated, is named and bounded so that the reader can judge the residual risk rather than discover it.

### 5.5.1 Internal validity

The chief internal-validity threat is the autonomy-hardware collinearity treated in Section 5.4.4, mitigated by the within-rover prong and confronted by the between-rover prong. The second internal threat is bad controls, mitigated by the pre-committed mediator rule of Section 5.4.2 that keeps realized slip and realized duration on the outcome side. The third internal threat is terrain selection by mission: planners route each rover along paths chosen partly for drivability, so observed terrain is endogenous to the rover's capability, and a more capable rover may have been sent across terrain that an earlier rover would have avoided. The mitigation is the terrain-class fixed effect and the within-terrain-class comparison, which restrict the contrast to drives on comparable ground; the residual risk is that terrain class is a coarse partition and within-class terrain may still differ systematically with rover, which is conceded as a caveat rather than claimed to be solved. A fourth internal threat is operational learning by the ground team over the twenty-year program, which could raise productivity independently of onboard autonomy. The mitigation is the autonomous-drive-fraction measure itself: ground-team learning improves blind commanded drives as well as autonomous ones, so it raises productivity across the board within a rover and is absorbed by neither the autonomous-fraction variation nor cleanly by the rover fixed effect, but the within-rover contrast between autonomous and blind segments on the same machine in the same operational era isolates the onboard contribution from the team's general improvement. This mitigation is partial, because a team that has learned to plan autonomous drives well is improving the autonomy channel specifically; the design treats that as part of the autonomy effect rather than a confounder, on the ground that the operational realization of onboard autonomy is inseparable from, and properly attributed to, the autonomy capability.

### 5.5.2 External validity
The external-validity threat is that the fleet is four rovers at three landing sites on one planet, so generalization beyond the Mars surface program is not licensed by the estimation. A result on this fleet does not transfer automatically to lunar rovers, to future Mars rovers with different flight processors, or to off-road terrestrial autonomy, and the design makes no such claim. The mitigation is to state the boundary of generalization explicitly and to separate the empirical claim about this fleet from the structured conjecture about transfer. The Mokyr lens supplies the conjecture: the autonomy channel should generalize better than the hardware channel, because the propositional base underlying perception and planning transfers across platforms while a specific wheel does not (Mokyr [\[85\]](#ref-mokyr2002)). That conjecture is labeled a hypothesis for future replication rather than a result of this design, which is the honest disposition of an external-validity threat that the data cannot resolve.

### 5.5.3 Construct validity

The construct-validity threats concern whether the measured quantities capture the constructs they are meant to capture. Meters per sol is a defensible but partial measure of mobility productivity. It counts distance and time but ignores the scientific value of where the rover went; a rover that drives far across barren ground scores high without advancing the mission. The mitigation is to be explicit that the construct measured is traverse productivity, not science return, and to note in the discussion that a confirmed autonomy effect on traverse productivity understates the operational value of autonomy, which also frees planning-cycle time (Farley et al. [\[45\]](#ref-farley2023); Gao and Chien [\[50\]](#ref-gao2021)). The second construct threat is that autonomy generation is a coarse three-level categorical that may not capture meaningful within-generation variation in autonomy capability. The mitigation is that the continuous autonomous-drive fraction is the finer and preferred treatment measure, with the categorical generation used mainly for the between-rover narrative; the autonomous fraction varies continuously and within rover, and thus captures the dimension the categorical misses. The third construct threat is that the generation labels might be mission self-description rather than genuine capability differences. The mitigation is the TechPort technology-readiness-level robustness check of Section 5.6.1, which classifies the maturity of each autonomy technology from an independent, mission-external record so the generation classification does not rest only on mission marketing.

### 5.5.4 Statistical-conclusion validity

The statistical-conclusion threat is that with few rover clusters, conventional cluster-robust standard errors are unreliable and will understate uncertainty, inflating the apparent significance of the between-rover contrasts. The mitigation is to compute inference for the between-rover contrasts by the wild-cluster bootstrap, the established remedy for few-cluster settings, and to report the few-cluster caveat wherever a between-rover coefficient is presented (de Chaisemartin and D'Haultfoeuille [\[28\]](#ref-chaisemartin2019); Goodman-Bacon [\[53\]](#ref-goodmanbacon2018)). A related threat is that the treatment effect is heterogeneous across terrain classes, so that a single pooled coefficient would be a weighted average with the potentially perverse weighting the recent literature warns of (de Chaisemartin and D'Haultfoeuille [\[29\]](#ref-chaisemartin2020); Goodman-Bacon [\[54\]](#ref-goodmanbacon2021); Callaway and Sant'Anna [\[23\]](#ref-callawaysantanna2021)). The mitigation is to estimate the heterogeneity directly by interacting autonomy with terrain class, reporting the effect by terrain class rather than forcing one homogeneous number. This respects the heterogeneity and turns it into a testable prediction: the autonomy effect should be largest in rough, hazard-dense terrain. A third statistical threat is multiple testing across the two outcome forms, the nested forms, and the terrain interactions. The mitigation is the pre-registration of Section 5.7, which fixes the primary test (the within-rover autonomous-fraction coefficient) in advance so that the confirmatory inference rests on a single pre-specified quantity and the remaining estimates are labeled exploratory.

## 5.6 The robustness battery

The robustness battery is the set of checks that probe whether the headline result survives reasonable alternative choices. Each check is specified in advance, so that it functions as a pre-committed stress test rather than a post-hoc search for a specification that confirms the hypothesis. The organizing claim is that a result that survives this battery is credible to the degree the battery is demanding, and the battery is designed to be demanding on exactly the dimensions where the design is most vulnerable.

### 5.6.1 TechPort generation re-labeling

The first check addresses the construct-validity worry that the autonomy-generation labels are mission self-description. The autonomy block is re-estimated with the generation classification replaced by the TechPort technology-readiness-level history of the mobility and autonomy technologies, an independent and mission-external record of when each capability reached flight maturity. If the autonomy effect is robust to this re-labeling, the effect is not an artifact of how the missions branded their own software; if the effect appears only with the mission labels and vanishes with the TechPort labels, the construct is suspect. This check is reported as a robustness column alongside the primary specification, with the TechPort record identifiers documented so the classification is auditable.

### 5.6.2 Terrain-interaction heterogeneity

The second check is the terrain-interaction estimation already motivated under statistical-conclusion validity, here cast as a robustness and mechanism check. The autonomous-fraction coefficient is allowed to vary by terrain class, and the pattern is examined against the mechanism prediction. The mechanism, named in full, runs as follows: the driver is onboard hazard detection and real-time path planning; the rover plans around hazards during the drive rather than waiting for a ground decision; the observable effect is that the autonomy advantage concentrates where hazards are dense and blind driving is most penalized; the operational consequence is more meters per sol in rough terrain specifically; the strategic implication is that the value of onboard autonomy rises with terrain difficulty (Verma et al. [\[112\]](#ref-verma2025); Arvidson et al. [\[18\]](#ref-arvidson2017)). If the estimated interaction is flat, with no concentration of the autonomy effect in rough terrain, the mechanism is unsupported even if a pooled correlation exists, and the design treats a flat interaction as one of the three falsification conditions. The check thus does double duty: it probes robustness and it tests the proposed mechanism, which is stronger than testing the average effect alone.

### 5.6.3 Drive-level secondary unit

The third check re-runs the analysis at the individual-drive level rather than the drive-sol level, wherever the PDS archive resolves multiple drives within a sol. The drive-sol is the primary unit because it aligns with the planning cycle and with the way the archive most consistently reports, but aggregating to the sol could mask within-sol variation or introduce aggregation artifacts. If the autonomous-fraction effect is stable across the drive-sol and the drive units, the result is not an artifact of the aggregation choice; if it changes, the aggregation sensitivity is reported. This check is bounded by data availability, because not every sol resolves to multiple archived drives, and the design states that limitation rather than promising a complete drive-level panel.

### 5.6.4 Alternative functional form and outlier sensitivity

The fourth check addresses functional form and influential observations. Because the panel is small at the cluster level, a single rover or a single extreme drive-sol (for example a record-setting long autonomous drive) could exert outsized leverage. The mitigation is to re-estimate with each rover dropped in turn (leave-one-rover-out), to re-estimate trimming the most extreme drive-sols, and to compare the linear specification against a specification in logarithms of distance and sols, which is natural given the multiplicative structure of the meters-per-sol-versus-sols-per-meter pair. A result that survives leave-one-rover-out and trimming is not driven by a single influential unit; a result that depends on one rover or a handful of drives is flagged as fragile. This check matters here because, with only three full panel members, leave-one-rover-out is a stringent test that removes a third of the between-rover information at a time.

### 5.6.5 Measurement-error sensitivity on the autonomous fraction

The fifth check addresses the moderate evidence gap flagged in the design specification: a clean published per-drive autonomous-fraction series across all generations is not guaranteed in the public record, so the fraction is partly reconstructed from the performance reports rather than read off directly (Verma et al. [\[112\]](#ref-verma2025); Maimone, Cheng, and Matthies [\[79\]](#ref-maimone2007); Biesiadecki and Maimone [\[20\]](#ref-biesiadecki2007)). Classical measurement error in a right-hand-side variable biases its coefficient toward zero, which would work against H1. The mitigation is to bound the reconstruction: the autonomous fraction is constructed under a high-autonomy and a low-autonomy reading of the ambiguous segments, and the coefficient is reported under both, so the sensitivity of the result to the reconstruction is visible. Where the corpus lacks the direct measurement, the design states the gap and reports the bounded sensitivity rather than presenting a single reconstructed series as if it were measured without error. This is the highest-value gap to close before execution, and the robustness battery treats it as such rather than papering over it.

## 5.7 Power and minimum-detectable-effect analysis

### 5.7.1 The two-level structure of power in this design

A power analysis asks how large an effect the design can detect at a given significance level and power, and the honest answer here has two parts because the design has two levels of variation. The design is well powered at the within-rover, drive-sol level and weakly powered at the between-rover level, and this asymmetry is exactly why the within-rover contrast is the primary test and the between-rover contrast is secondary. The reason lies in the structure of the panel. The within-rover analysis has hundreds to thousands of drive-sols per rover (the Mars Exploration Rovers and Curiosity each accumulated thousands of sols of operation, and Perseverance hundreds at the time of a given archive cut), while the between-rover analysis has effectively three independent clusters. This follows from the standard result that the precision of a within-unit slope is governed by the number of within-unit observations and the residual variance, while the precision of a between-unit contrast is governed by the number of units, which here is small (Angrist and Pischke [\[7\]](#ref-angrist2009); de Chaisemartin and D'Haultfoeuille [\[28\]](#ref-chaisemartin2019)).

### 5.7.2 Minimum detectable effect at the within-rover level

The minimum detectable effect (MDE) for the within-rover autonomous-fraction coefficient is sized as follows, with every input labeled an assumption pending the assembled panel. The MDE for a regression slope at significance level alpha and power 1 minus beta is approximately the product of the relevant critical-value sum and the standard error of the slope, where that standard error scales inversely with the square root of the effective number of within-rover observations and the standard deviation of the autonomous fraction, and directly with the residual standard deviation of productivity. Take an illustrative residual standard deviation of meters per sol on the order of the cross-drive spread reported informally in the engineering literature, an illustrative within-rover standard deviation of the autonomous fraction of roughly a quarter (reflecting that the fraction ranges widely across segments), and an effective sample on the order of one thousand drive-sols per rover after accounting for within-rover serial correlation. Under these inputs the design can detect a slope corresponding to a change of a few meters per sol per unit change in autonomous fraction at conventional five-percent significance and eighty-percent power. These figures are illustrative, computed to size the design rather than estimated from data, and they establish that the within-rover test is not underpowered for effects of the magnitude the mechanism predicts: the Enhanced Navigation step is described as a substantial increase in achievable distance per sol, far larger than the MDE just sketched (Verma et al. [\[112\]](#ref-verma2025)). The conclusion is conditional and calibrated. If the true within-rover autonomy effect is anywhere near the magnitude the qualitative literature suggests, the design has ample power to detect it; if the true effect is small, the design may miss it, and a null within-rover result would then be ambiguous between a true small effect and insufficient power.

### 5.7.3 The between-rover power deficit and its honest disposition

At the between-rover level the design is candidly underpowered, and saying so is part of its credibility rather than a weakness to be hidden. With three effective clusters, the wild-cluster bootstrap that delivers honest inference also delivers wide confidence intervals, and the MDE for a whole-generation contrast is correspondingly large. The disposition of this deficit is threefold. First, the between-rover contrast is explicitly secondary and is never the basis for a confirmatory claim about H1. Second, the few-cluster caveat accompanies every between-rover number, so a reader cannot mistake a wide interval for a precise zero. Third, the design does not attempt to manufacture power it does not have by treating drive-sols as if they were independent observations at the rover level, the cardinal error the few-cluster literature warns against (de Chaisemartin and D'Haultfoeuille [\[28\]](#ref-chaisemartin2019); Goodman-Bacon [\[53\]](#ref-goodmanbacon2018)). The power analysis thus ends not with a single number but with a calibrated statement: high power for the primary within-rover test, low power for the secondary between-rover test, and a confirmatory burden placed entirely on the level where power is adequate.

## 5.8 The pre-registration commitment
### 5.8.1 What is pre-registered and why

The design commits to pre-registration, which here is not a bureaucratic formality but the mechanism that makes the contribution falsifiable in the strong sense the prospectus requires. A design with two outcome forms, three nested specifications, terrain interactions, and a five-part robustness battery carries enough researcher degrees of freedom that, absent a pre-committed primary test and decision rule, a determined analyst could find some specification to confirm almost any hypothesis. The governing principle comes from the credibility revolution: the value of a result depends on the transparency and pre-commitment of the analysis that produced it, not on the result's agreement with prior belief (Angrist and Pischke [\[10\]](#ref-angristpischke2010); Leamer [\[72\]](#ref-leamer2010)). It reflects the broad movement in applied econometrics toward specifying the analysis before seeing the outcome data, so that confirmatory and exploratory claims are distinguished by their timing relative to the data rather than asserted after the fact.

The pre-registration fixes the following items before any productivity outcome is merged into the panel. The primary hypothesis test is the sign and significance of the within-rover autonomous-fraction coefficient \(\beta_{\text{af}}\) in the meters-per-sol specification, with the sols-per-meter specification as a pre-specified parallel. The primary inference is conditional mean independence after the rover and terrain fixed effects and the terrain block, with wild-cluster bootstrap inference reserved for the secondary between-rover contrasts. The covariate set is fixed: a priori terrain and commanded-drive parameters are eligible; all realized quantities are mediators and excluded. The decision rule is fixed and is restated in Section 5.8.2. The robustness battery of Section 5.6 is named in full, so that the checks function as confirmatory stress tests rather than a menu from which a favorable subset is later selected. Anything not pre-registered, for example a terrain partition discovered to be informative only after exploration, is labeled exploratory in the eventual reporting.

### 5.8.2 The fixed decision rule

The decision rule on H1 against H0 is committed in advance and stated here so that Chapter 6 can operate on it without renegotiation. H1 is supported if and only if the within-rover autonomous-fraction coefficient \(\beta_{\text{af}}\) is positive and significant in the meters-per-sol specification under the pre-registered inference, and survives the leave-one-rover-out and measurement-error-bounding robustness checks. H1 is rejected and H0 stands if \(\beta_{\text{af}}\) is indistinguishable from zero within rover after the terrain conditioning, or if the autonomy-generation block loses joint significance and incremental explanatory share the moment the hardware block enters the between-rover nested decomposition, or if the terrain-interaction pattern is flat so that the proposed hazard-avoidance mechanism is unsupported. These are the three falsification conditions, named before estimation, and their pre-commitment is what distinguishes a falsifiable contribution from an illustrated one (Angrist and Pischke [\[7\]](#ref-angrist2009)). A result that satisfies the primary test but fails a robustness check is reported as fragile rather than as either a clean confirmation or a clean rejection, and the design does not collapse that intermediate case into one of the two poles for narrative convenience.

### 5.8.3 The order of operations that protects the pre-registration

A pre-registration is only as good as the order in which the data are touched, so the design fixes that order. The panel is assembled and the treatment, covariates, and fixed effects are constructed and validated against known values before the productivity outcome is merged. The measurement decisions (terrain-class crosswalk, autonomous-fraction reconstruction with its high and low bounds, hardware-covariate construction) are finalized on the treatment-and-covariate side first. Only then is the outcome merged and the pre-registered estimation run. This sequencing prevents the outcome from influencing the construction choices, which is the channel through which pre-registrations most often leak. The order of operations is documented as part of the computational plan in Section 5.9 and is auditable from the version-control history of the analysis code.

## 5.9 The computational and software plan

### 5.9.1 Reproducibility as a design requirement

The computational plan treats reproducibility as a requirement of the design rather than an afterthought, because a design argued to the standard of a pre-registration is credible only if the analysis that eventually executes it is fully reproducible from public data and version-controlled code. All three data sources are public (the PDS Geosciences Node archives, the NASA Technical Reports Server, and NASA TechPort), so the entire pipeline from raw archive to estimated coefficient can in principle be reconstructed by an independent party. This answers the transparency standard of the credibility movement, under which a result that cannot be reproduced from stated inputs and stated code is not a result others can build on (Angrist and Pischke [\[10\]](#ref-angristpischke2010); Leamer [\[72\]](#ref-leamer2010)). The plan therefore specifies the pipeline in enough detail that the next phase of work, the execution, has no undocumented discretion.

### 5.9.2 The pipeline

The pipeline has five stages, each a separately version-controlled and separately testable module. Stage one ingests the PDS traverse and localization products for the Mars Exploration Rovers, Curiosity, and Perseverance, parsing per-drive localized path length and elapsed sols and constructing the drive-sol panel, with Sojourner handled as a boundary case rather than a full panel member because its archive predates the traverse-product standard. Stage two merges the autonomy-generation classification and the autonomous-drive fraction from the NTRS performance reports, building the fraction under the high and low reconstruction bounds of Section 5.6.5. Stage three merges the hardware covariates from published mission descriptions and the TechPort records, and constructs the terrain class and terrain covariates from orbital basemaps and PDS terrain characterizations (Golombek et al. [\[52\]](#ref-golombek2014); Arvidson et al. [\[18\]](#ref-arvidson2017)). Stage four runs the estimation: the two outcome forms, the three nested between-rover forms, the primary within-rover specification, and the terrain interactions, with the fixed effects absorbed by an appropriate high-dimensional fixed-effects routine. Stage five computes inference, the wild-cluster bootstrap for the between-rover contrasts and conventional within-rover standard errors clustered to respect serial correlation within rover, and runs the robustness battery as a fixed sequence of additional estimations. The separation into stages is deliberate: the outcome merge happens at stage four and not before, which enforces the order-of-operations protection of Section 5.8.3.

### 5.9.3 Software and statistical tooling

The statistical computation uses an open-source environment so the toolchain itself is reproducible. The high-dimensional fixed-effects estimation, which must absorb rover and terrain-class fixed effects efficiently, is performed with an established fixed-effects regression implementation; the wild-cluster bootstrap is performed with an established few-cluster bootstrap implementation; and the data assembly is scripted rather than performed interactively so that every transformation is recorded. The point is methodological rather than a software endorsement: the design requires that the fixed-effects absorption and the few-cluster bootstrap be done by validated, documented routines whose behavior in the few-cluster regime is understood, because a hand-rolled implementation of either is exactly where a small-sample design most easily goes wrong (de Chaisemartin and D'Haultfoeuille [\[28\]](#ref-chaisemartin2019); Goodman-Bacon [\[53\]](#ref-goodmanbacon2018)). Random seeds for the bootstrap are fixed and recorded so the inference is bit-for-bit reproducible. The full set of access paths, query strings, and record identifiers for the three archives is documented in the appendices so that an independent party can re-ingest the raw data, and the analysis code is structured so that re-running it from the public archives reproduces every number in the eventual results chapter.

### 5.9.4 Validation and quality control

Before any estimation, the constructed panel is validated against known values: total traverse distances per rover are checked against the published cumulative odometry figures for the Mars Exploration Rovers, Curiosity, and Perseverance, and discrepancies beyond a documented tolerance are investigated rather than ignored (Maimone, Cheng, and Matthies [\[79\]](#ref-maimone2007); Farley et al. [\[45\]](#ref-farley2023)). The autonomous-fraction reconstruction is sanity-checked against the qualitative descriptions in the performance reports, so that, for example, Perseverance's Enhanced Navigation era shows a higher reconstructed fraction than the Mars Exploration Rover era, consistent with the documented capability step (Verma et al. [\[112\]](#ref-verma2025); Di et al. [\[41\]](#ref-di2022); Maki et al. [\[80\]](#ref-maki2020)). The terrain-class crosswalk is checked against the terrain characterizations in the literature so that drives through known megaripple and dune fields are classified as rough rather than benign (Arvidson et al. [\[18\]](#ref-arvidson2017); Golombek et al. [\[52\]](#ref-golombek2014)). These validation steps are part of the design, not optional hygiene, because a design whose primary identification rests on within-rover terrain conditioning is only as good as the terrain classification that does the conditioning. Where a validation check cannot be satisfied because the public record is thin, as the design specification flags for the autonomous-fraction series and the terrain crosswalk, the gap is documented and the affected result is reported with the corresponding sensitivity bound rather than presented as if the validation had passed.

## 5.10 Chapter synthesis: the design and its calibrated confidence

The chapter closes by drawing the threads together, with the confidence in each claim stated rather than asserted. The productivity rise and the combined-narrative description of it are documented across the mission literature, and the absence of any prior identification that separates the channels was established in Chapter 3, so the reality of the problem is held at high confidence. Its materiality is held at high confidence too: sols are the binding scarcity for the campaign the fleet now serves, which makes the choice of the channel that buys productivity most cheaply a decision of direct cost and risk consequence. The claim that the design reaches the causal mechanism carries moderate-to-high confidence. The within-rover autonomous-fraction contrast holds hardware fixed at the level of the machine and, under the conditional mean independence argued in Section 5.4, recovers the autonomy effect, with the bad-controls rule protecting the estimand and the signed-direction argument making a positive result conservative; the confidence is moderate rather than high because conditional mean independence is an assumption the design makes defensible but cannot prove, and because the autonomous-fraction measurement is partly reconstructed. That the design improves on its rivals is held at high confidence: no instrument is available, a staggered difference-in-differences is the wrong frame for a continuous within-rover treatment, and the within-rover panel is the cleanest comparison the fleet permits, with the between-rover prong bounding the broader question. The remaining identification risk is held at moderate confidence: the few-cluster inference, the leave-one-rover-out checks, the measurement-error bounds, and the pre-registered decision rule make it transparent and bounded, while the small cross-section and the terrain endogeneity remain genuine, conceded limitations that no design over this fleet can eliminate.

What would raise these confidence grades is concrete and is named so the next phase has a target: a clean, published per-drive autonomous-fraction series across all generations would raise the design grade from moderate toward high by removing the measurement-error caveat; a richer terrain partition validated against in-situ rather than only orbital data would tighten the terrain conditioning that the within-rover identification leans on; and a fourth full panel member, were a future rover to enter the archive, would relax the between-rover power deficit. What would lower the grades is equally concrete: evidence that the ground team's autonomy decisions are driven by within-terrain-class factors the terrain block does not capture would weaken conditional mean independence, and evidence that the generation labels do not survive the TechPort re-labeling would weaken the construct. The design is honest about both directions, which is the disposition a falsifiable contribution requires.

The estimator, the specifications, the identification argument, the threat treatment, the robustness battery, the power analysis, the pre-registration, and the computational plan together constitute the design-stage contribution. They convert the autonomy-versus-hardware question into a single pre-registered test with a fixed decision rule, an honest account of what it can and cannot detect, and a reproducible path from public archive to coefficient. Chapter 6 takes this design as given and specifies the analysis plan and the labeled, illustrative expectations that follow from the named causal mechanism, without executing the estimation that the next phase of work will perform.



# Chapter 6: Analysis Plan and Expected Results

## 6.1 Chapter thesis and what this chapter commits to in advance

The analysis is pre-registered with a fixed decision rule on H1 against H0, the expected signs of every coefficient follow deductively from the named causal mechanism rather than from any look at the data, and every number that appears in this chapter is an expectation or an illustration, never an executed estimate. That is the chapter thesis, and it governs everything that follows. The purpose of a pre-registered analysis plan is to bind the analyst's hands before the panel is assembled, so that the verdict on the hypothesis cannot be steered after the fact by the choice of specification, the choice of controls, or the choice of which result to foreground. The discipline this chapter enforces is the one that distinguishes a falsifiable empirical claim from an illustrated one: the design states, in advance and in writing, exactly what pattern of estimates would force the conclusion that H1 is wrong (Angrist and Pischke [\[7\]](#ref-angrist2009)).

Precision about the current state of this work, the desired state, the gap between them, and the consequence of leaving the gap open is what makes the analysis plan a contribution in its own right rather than a placeholder for results not yet obtained. The current state is that the prior chapters have fixed the research question, the hypotheses, the variable definitions, the estimator, and the identification strategy. What is not yet fixed, and what this chapter fixes, is the exact sequence of estimation steps, the exact decision rule that maps the resulting estimates onto a verdict, and the exact set of falsification checks whose failure would reject the contribution. The desired state is a plan complete enough that a second analyst, handed the same three archives, would execute the same steps, apply the same decision rule, and reach the same verdict, with no residual researcher degrees of freedom that could swing the answer. The gap is that without such a plan, the temptation in a small-sample panel with collinear regressors is overwhelming: one can almost always find a specification in which the autonomy block looks decisive, and almost always find another in which the hardware block does, because with three full panel members and two correlated rover-level treatments the data underdetermine the answer unless the rule for reading them is fixed first. The consequence of leaving the gap open is that the eventual estimates, whatever they are, would carry no evidentiary weight, because a reader could not distinguish a genuine finding from a specification chosen to produce it. Closing the gap is therefore not a formality. It is the step that converts the design of Chapter 5 into a test that can actually fail.
This chapter remains, like its siblings, a design-stage document. No panel has been assembled, no regression has been run, and no coefficient reported here is an estimate. Where the chapter states an expected sign or an illustrative magnitude, it does so to make the design concrete and to show that the decision rule is well defined over the space of possible results, not to report a finding. Every such value is labeled illustrative or expected at the point of use, and every result table specified in this chapter is left deliberately unpopulated, its columns and rows named and its cells empty by design. The honesty commitment from the prospectus and the design specification carries here without exception: the contribution at this stage is the design, and the analysis plan is the part of the design that says what will be done with the data and how the answer will be read once the data speak.

The chapter proceeds in five movements that follow the logical order of the work. Section 6.2 lays out the estimation procedure, from panel assembly through the nested decomposition, the within-rover test, the few-cluster bootstrap, and the falsification battery. Section 6.3 states the fixed decision rule on the hypothesis, the single most important commitment in the chapter, because it is the rule that cannot be revised after the estimates are seen. Section 6.4 derives the expected sign and relative magnitude of each coefficient from the causal mechanism, labeling every value illustrative and showing that the expectations are mechanism-driven predictions rather than hopes. Section 6.5 specifies the three falsification checks in operational detail. Section 6.6 designs the illustrative simulation that demonstrates the estimator recovers a known effect under a known data-generating process, and specifies the descriptive generation profile and autonomy-fraction-profile interpretations, with all result tables specified but unpopulated. A short closing section draws the chapter together and states its calibrated confidence.

## 6.2 The step-by-step estimation procedure

The estimation procedure is a fixed sequence. Its order is not arbitrary: each step produces an input required by the next, and the falsification checks come last so that they cannot be used to tune the specifications that precede them. The procedure reproduces and extends the seven-step skeleton from the prospectus and from Chapter 5, with each step elaborated into the operational detail a second analyst would need.

### 6.2.1 Step one: assemble the drive-sol panel

The first step assembles the drive-sol panel from the Planetary Data System traverse and localization products for the three full panel members, the two Mars Exploration Rovers, Curiosity, and Perseverance. For each rover, each sol on which a drive occurred, and each commanded drive within that sol, the procedure extracts the localized traverse path and computes its arc length, which is the localized path length rather than the straight-line displacement between start and end points. This distinction is load-bearing. Two drives with the same net displacement can differ greatly in path length when one of them detours around hazards, and the path-length measure is the one that captures the actual mobility work the rover performed. The drive-level records are then aggregated to the drive-sol, summing localized path length across all drives within the sol and recording the elapsed sols, so that the primary unit of analysis is one rover, one sol, with one or more drives collapsed onto it. The panel is unbalanced by construction, because rovers differ in mission duration and in the density of drive sols across their records, and the procedure carries the unbalancedness rather than forcing a balanced design that would discard usable observations. Sojourner is not entered as a fourth panel member: its traverse record predates the PDS traverse-product standard, the public archive carries almost no drive-level Sojourner product, and the prospectus designates it a boundary case to be treated qualitatively in Chapter 4 rather than a full panel member. This analysis plan honors that designation by excluding Sojourner from the estimating sample and noting its absence as a coverage limit rather than papering over it with fabricated observations.

The localization that ties each drive's path to a consistent spatial frame is itself a measured quantity with its own provenance, drawn from the rover-localization-within-orbital-maps methodology that registers surface position against orbital basemaps. The procedure records, for each drive-sol, the localization method and any flagged localization uncertainty, so that drive-sols with degraded localization can be down-weighted or excluded in a robustness pass rather than silently entering the panel at full weight. The output of step one is a clean drive-sol panel keyed by rover and sol, carrying localized meters traversed and elapsed sols, which together construct both forms of the dependent variable: meters per sol directly, and sols per meter as its inverse.

### 6.2.2 Step two: merge the autonomy and hardware blocks

The second step merges, onto each drive-sol, the autonomy-generation classification, the continuous autonomous-drive fraction, the hardware covariates, and the TechPort technology-readiness records. The autonomy-generation indicator is categorical with three levels, G1 for the Mars Exploration Rover-class AutoNav with visual odometry, G2 for the Mars Science Laboratory inherited and extended stack, and G3 for the Mars 2020 Enhanced Navigation, and it is assigned at the rover level because a given rover flies a given generation throughout its mission. The continuous autonomous-drive fraction, by contrast, is assigned at the drive-sol level, because the same rover executes some drive segments under onboard autonomous navigation and others under blind commanded motion depending on terrain and ground decisions, and it is this within-rover variation in the fraction that carries the primary identification of H1. The fraction is reconstructed from the NTRS AutoNav and Enhanced Navigation performance reports (Maimone, Cheng, and Matthies [\[79\]](#ref-maimone2007); Biesiadecki and Maimone [\[20\]](#ref-biesiadecki2007); Verma et al. [\[112\]](#ref-verma2025)), and the procedure records, for each drive-sol, both the reconstructed fraction and a flag indicating the directness of its source. The design specification flags the per-sol autonomous-fraction series as the single highest-value evidence gap, so the analysis plan must carry an explicit measurement-error treatment for the fraction rather than pretending it is observed without error.

The hardware covariates, wheel diameter, mass class, actuator class, and nominal available drive energy per sol, are merged at the rover level from the published mission descriptions and the TechPort records. Because these covariates are rover-level and time-invariant within a rover, they are collinear with rover identity and with autonomy generation, which is the central identification problem named in Chapter 5 and resolved by the two-pronged design. The TechPort technology-readiness records are merged as an independent, mission-external classification of the maturity of each autonomy and mobility technology at the time each rover was designed, and they are carried not as a primary regressor but as a robustness input: they are program records rather than peer-reviewed literature, so the design specification correctly names them as a dataset rather than as a corpus citation, and the analysis plan uses them only to check that the G1/G2/G3 labels reflect genuinely distinct technology generations rather than mission self-description.

### 6.2.3 Step three: merge the terrain block

The third step merges terrain class and the continuous terrain covariates onto each drive-sol. Terrain class is the categorical dimension that the fixed-effects design absorbs, and the continuous terrain block, slope and a physical-properties index, varies within terrain class and enters as a covariate. The terrain assignment is constructed from orbital basemaps and from the PDS-archived terrain characterizations and terrain-physical-properties derivations (Golombek et al. [\[52\]](#ref-golombek2014); Arvidson et al. [\[18\]](#ref-arvidson2017)), and the procedure records, for each drive-sol, the a priori terrain class and slope as they would have been known to the ground team at planning time, not the terrain as revealed by the drive itself. This timing matters for the bad-controls discipline: the terrain block is an eligible control precisely because it is a priori, fixed before the drive, and therefore not a consequence of how the autonomy software chose to drive, whereas anything realized during the drive is a mediator and stays off the right-hand side. The design specification flags the terrain-class-by-drive crosswalk as a construction step performed by the candidate rather than a ready-made cited table, so the procedure documents the crosswalk rules explicitly and treats the crosswalk as a documented data-construction artifact subject to its own sensitivity analysis, not as an unquestioned input.

The output of steps one through three is the analysis-ready panel: one row per drive-sol, carrying the two forms of the dependent variable, the autonomy-generation indicator, the autonomous-drive fraction with its source flag, the hardware block, the terrain class, and the continuous terrain covariates, with realized slip and realized wheel-soil interaction recorded but quarantined on the outcome side as mediators that never enter the estimating equations.

### 6.2.4 Step four: the nested decomposition

The fourth step estimates the nested decomposition that answers the between-rover form of the question. Three specifications are estimated in sequence, each adding one block of regressors, for both the meters-per-sol and the sols-per-meter forms of the dependent variable. The first nested form includes only the terrain-class fixed effects and the continuous terrain covariates, establishing how much of the variation in productivity is accounted for by terrain alone. The second nested form adds the hardware block, measuring the incremental explanatory share attributable to mechanical platform differences once terrain is held fixed. The third nested form adds the autonomy block, measuring the incremental explanatory share attributable to autonomy generation once both terrain and hardware are held fixed. The quantity of interest from this decomposition is the comparison of the incremental explanatory share contributed by the autonomy block against that contributed by the hardware block, together with the behavior of the hardware coefficients as the autonomy block enters: if part of what looked mechanical was in fact the correlated autonomy upgrade, the hardware coefficients should attenuate when the autonomy block is added. The procedure records, for each nested form, the explanatory share and the coefficient block, so that the decomposition can be read as a sequence rather than as a single regression, because it is the sequence, not any one specification, that reveals how the two collinear channels partition the variance.

The notation for every specification in this step is the fixed notation from the design specification and from Chapter 5, reproduced here exactly so that no drift creeps in between chapters. For drive-sol *i* on rover *r* in terrain class *c*:

\[
\text{Productivity}_{irc} = \beta_1 \, \text{AutonomyGen}_{r}
                 + \beta_2 \, \text{Hardware}_{r}
                 + \gamma \, \text{Terrain}_{ic}
                 + \alpha_{r} \quad \text{(rover fixed effect)}
                 + \delta_{c} \quad \text{(terrain-class fixed effect)}
                 + \epsilon_{irc} \qquad\qquad (1)
\]

In the nested between-rover decomposition the rover fixed effect \(\alpha_{r}\) is relaxed to the terrain-class fixed effect \(\delta_{c}\) plus the hardware block, because a rover fixed effect that fully absorbed rover identity would also absorb autonomy generation, which is a rover-level attribute, and would leave the generation coefficient unidentified. This is the deliberate tension Chapter 5 confronts: the between-rover contrast identifies the generation effect by leaning on the hardware block to separate the channels, and it is reported as a complement to, not a substitute for, the within-rover contrast that follows.

### 6.2.5 Step five: the within-rover autonomous-fraction estimation

The fifth step estimates the within-rover specification that is the primary test of H1. Here the rover fixed effect \(\alpha_{r}\) is restored to its full strength, absorbing all fixed, rover-specific characteristics including the entire hardware block, and the regressor of interest is the continuous autonomous-drive fraction, which varies within rover across drive-sols. The terrain-class fixed effects and the continuous terrain covariates remain in the specification, because the autonomous fraction is chosen by the ground team rather than assigned at random, and that choice is correlated with terrain: planners drive a segment autonomously when they judge the terrain suitable and the time pressure high, and blind when the terrain is benign or the route short and well-imaged. Conditioning on a priori terrain is therefore not an optional refinement but the core of the identification, because it conditions out the terrain-driven component of the autonomous-fraction choice so that the residual variation in the fraction is closer to as-good-as-random with respect to productivity. The coefficient on the autonomous fraction in this specification is the closest thing in this setting to the design-based ideal of comparing like with like, because within a single machine the hardware is constant by construction and the only thing differing between a more-autonomous and a less-autonomous segment is how much the software is doing (Angrist and Pischke [\[7\]](#ref-angrist2009); Angrist and Pischke [\[10\]](#ref-angristpischke2010)).

Because the autonomous fraction is reconstructed rather than directly observed for every drive-sol, this step carries an explicit measurement-error treatment. Classical measurement error in a right-hand-side regressor attenuates its coefficient toward zero, which would bias the within-rover test against H1 and in favor of the null, so the procedure treats a non-trivial within-rover autonomous-fraction coefficient as conservative evidence and treats a null as potentially attenuation-driven rather than as clean support for H0. Where the directness flag from step two indicates a more direct source for the fraction, the procedure estimates the specification on the higher-confidence subset as a robustness pass, and the comparison of the full-sample and high-confidence-subset coefficients is itself informative about the magnitude of attenuation.

### 6.2.6 Step six: few-cluster-aware inference

The sixth step computes inference that respects the panel structure. With repeated drive-sols nested within a small number of rovers, the effective number of independent clusters is small, and conventional cluster-robust standard errors understate uncertainty and over-reject the null of no effect. The procedure therefore computes wild-cluster-bootstrap standard errors and confidence intervals for the between-rover contrasts, the contrasts most exposed to the few-cluster problem because they rest on differences across the three full panel members, and it reports the few-cluster caveat explicitly alongside every between-rover estimate. The within-rover autonomous-fraction contrast is less exposed, because its identifying variation is within-rover and the number of drive-sols within each rover is large, but the procedure nonetheless clusters at the rover level for the within-rover inference as well, on the conservative principle that the residuals are plausibly correlated within rover. The choice of wild-cluster bootstrap over conventional cluster-robust inference is dictated by the modern fixed-effects literature on estimation and inference with few and heterogeneous clusters, which establishes that the bootstrap is the appropriate tool when the cluster count is small and that conventional asymptotics are unreliable in that regime (de Chaisemartin and D'Haultfoeuille [\[29\]](#ref-chaisemartin2020); Goodman-Bacon [\[54\]](#ref-goodmanbacon2021)). The procedure also records, for the generation contrasts estimated across the staggered introduction of G1, G2, and G3, the diagnostics from the heterogeneous-and-dynamic-effects literature, because a naive two-way fixed-effects estimator can place negative weights on some treated-group comparisons when treatment effects are heterogeneous across groups and timing, and the procedure must verify that the generation contrast it reports is not a negatively weighted average of underlying effects (de Chaisemartin and D'Haultfoeuille [\[29\]](#ref-chaisemartin2020); Callaway and Sant'Anna [\[23\]](#ref-callawaysantanna2021); Goodman-Bacon [\[53\]](#ref-goodmanbacon2018)).

### 6.2.7 Step seven: the falsification battery

The seventh and final step runs the three falsification checks specified in Section 6.5. These come last in the sequence so that they cannot be used to tune the specifications that precede them. Each check is a pre-specified test whose failure rejects a component of the contribution, and the order of the battery is fixed: the within-rover null check first, because it is the most direct, the generation-collapse check second, and the terrain-interaction-flatness check third. The output of step seven is the verdict on H1 against H0, read off the fixed decision rule of Section 6.3, reported with the few-cluster-aware uncertainty from step six and accompanied by the honest statement of which falsification checks the data passed and which they failed.
## 6.3 The fixed decision rule on the hypothesis

The decision rule is the single most important commitment in this chapter, because it maps the eventual estimates onto a verdict and must be fixed before any estimate is seen. The rule is stated as a conjunction of conditions, all of which must hold for H1 to be sustained, and the failure of any one of which moves the verdict toward H0. Stating the rule as a conjunction rather than a disjunction is deliberate and conservative. It makes H1 harder to sustain, because it requires the contribution to survive every test rather than merely one, and it guards against the small-sample temptation to declare victory on the strength of a single favorable specification.

H1 is sustained if and only if all three of the following hold. First, in the within-rover specification of step five, the coefficient on the autonomous-drive fraction is positive in the meters-per-sol form, negative in the sols-per-meter form (these being the same sign requirement expressed on the two inverse measures), and distinguishable from zero under wild-cluster-bootstrap inference clustered at the rover level. This is the primary condition, because the within-rover contrast holds hardware fixed at the level of the individual machine and is therefore the cleanest available test of whether the software channel does any work at all. Second, in the nested decomposition of step four, the autonomy block contributes a larger incremental explanatory share than the hardware block when each is added last, and the hardware coefficients attenuate rather than strengthen when the autonomy block enters, which is the signature of the hardware block having absorbed correlated autonomy variation in its absence. Third, the terrain-interaction pattern of step five, estimated by interacting the autonomous fraction with terrain class, is not flat: the autonomy effect is larger in rougher and more hazard-dense terrain classes than on benign flats, which is the pattern the causal mechanism predicts and whose absence would undercut the mechanism even if a raw average effect were present.

H0 stands, and H1 is rejected, if any of the following occur. If the within-rover autonomous-fraction coefficient is indistinguishable from zero after terrain and rover fixed effects are included, the autonomy channel has no support at the level where hardware is held constant, and the contribution fails on its primary condition. If the autonomy-generation block loses joint significance and explanatory share the moment the hardware block enters the nested decomposition, the apparent autonomy effect is confounded with hardware and cannot be separated from it, and the contribution fails on its second condition. If the terrain-interaction pattern is flat, the proposed mechanism, that onboard hazard detection most reduces the blind-driving penalty where hazards are densest, is unsupported even should a raw correlation between autonomy and productivity exist, and the contribution fails on its third condition. The decision rule is therefore symmetric in its honesty: it specifies in advance not only what would sustain the contribution but exactly what would refute it, which is the property that makes the contribution falsifiable in the strong sense rather than merely illustrated (Angrist and Pischke [\[7\]](#ref-angrist2009)).

One refinement to the decision rule guards against a measurement-error false negative on the primary condition. Because the autonomous fraction is reconstructed and classical measurement error attenuates its coefficient toward zero, a within-rover coefficient that is small and statistically indistinguishable from zero in the full sample but positive and distinguishable from zero in the higher-confidence subset is read as weak support for H1 attenuated by measurement error, not as clean support for H0. This refinement is itself pre-specified, so that it cannot be invoked opportunistically after a disappointing full-sample result. The rule is that the high-confidence-subset estimate is consulted only when the full-sample estimate is null, and the verdict in that case is recorded as indeterminate-pending-better-measurement rather than as either sustaining or rejecting H1. This is the honest treatment of the single highest-value evidence gap the design specification identifies, and it converts a known data limitation into a stated caveat rather than a hidden bias.

## 6.4 Expected signs and the mechanism that generates them

This section derives the expected sign and the expected relative magnitude of each coefficient from the named causal mechanism. Every value in this section is illustrative or expected, labeled as such at the point of use, and none is an estimate produced from data. The point of the section is not to forecast numbers but to demonstrate that the expectations are deductions from a stated mechanism rather than free parameters, so that a reader can see the chain from driver to observable effect and can judge the prediction on its mechanism rather than on the analyst's optimism.

### 6.4.1 The proposed mechanism and a note on what the design identifies

A preliminary note on the scope of causal identification is necessary before the expected signs are derived, because the two estimators in the design do not have equal standing on this question. The within-rover autonomous-fraction contrast identifies a causal effect in the design-based sense: within a single machine, with the mechanical platform held fixed by construction, the only thing varying between a more-autonomous and a less-autonomous drive segment is how much of the perception and planning the software performed, and the terrain-class fixed effects condition out the primary driver of the autonomous-fraction choice. This comparison is the one the design calls causal, and it is the one the decision rule makes primary. The between-rover generation contrast, including the descriptive generation profile of Section 6.7.3, does not identify a causal effect: every new rover upgraded hardware and autonomy simultaneously, so the generation boundary is confounded with the platform upgrade, and no estimator in this design separates them. The generation contrast is described and analyzed throughout as associational and descriptive evidence, corroborating the within-rover result but not independently establishing a causal attribution.

With that boundary in place, the mechanism that generates the expected signs within the within-rover estimator can be stated. The proposed driver is the increase in the autonomous-drive fraction that accompanied the advance in autonomy-software generation from G1 through G2 to G3, culminating in Enhanced Navigation, which processes imagery and plans a path while the rover is still in motion rather than requiring it to stop, think, and then move (Verma et al. [\[112\]](#ref-verma2025)). The mechanism through which this driver acts on productivity is that a larger share of each drive is executed under onboard hazard detection and real-time path planning rather than under blind commanded motion, because a rover that can perceive and plan while driving need not pause for ground-in-the-loop decisions at every hazard. The observable effect is a higher autonomous-drive fraction and more localized meters per sol, concentrated in the hazard-dense terrain classes where the blind-driving penalty is largest. The operational consequence is fewer sols consumed per meter, which frees sols in the daily planning cycle for sampling, instrument placement, and caching. The strategic implication, which is the subject of Chapter 7 rather than of the estimation, is that surface mobility productivity may be bought more cheaply with software than with mass. Each expected sign below is pinned to a specific link in this mechanism, and where only a between-rover correlation rather than a within-rover identified comparison supports a sign, the section says so and downgrades the confidence in that expectation accordingly.

### 6.4.2 Expected sign of the within-rover autonomous-fraction coefficient

Illustrative, not estimated. The within-rover coefficient on the autonomous-drive fraction is expected to be positive in the meters-per-sol form and negative in the sols-per-meter form, and to survive the inclusion of rover and terrain-class fixed effects. The reasoning behind this expectation is the second link in the chain: within a single machine, with hardware constant by construction, the only thing differing between a more-autonomous and a less-autonomous segment of the same terrain class is how much of the perception and planning the software is doing, so a positive association between the autonomous fraction and meters per sol is the direct observable signature of the software channel doing real work. The reasoning is strong here because it rests on a within-unit comparison that holds the rival channel fixed, which is the design-based ideal rather than a bare correlation (Angrist and Pischke [\[10\]](#ref-angristpischke2010)). One qualification on this expectation is that the autonomous fraction is chosen by the ground team and is correlated with terrain, so the expectation holds only conditional on the a priori terrain block, and the objection the design must defend against is that residual terrain selection, not the software, drives the association; the terrain-class fixed effects and within-terrain-class comparison are the defense, and the residual-selection caveat is acknowledged as a genuine limit rather than dismissed. Confidence in this expectation is moderate to high at the design stage: high in the logic of the within-rover comparison, moderated downward by the measurement-error attenuation in the reconstructed fraction and by the residual-terrain-selection concern. What would raise confidence is a cleaner per-drive autonomous-fraction series from the performance reports; what would lower it is evidence that the autonomous-fraction choice tracks an unmeasured terrain feature that also drives productivity.

### 6.4.3 Expected pattern of the nested decomposition

Illustrative, not estimated. In the nested decomposition, the hardware block alone is expected to absorb a moderate share of between-rover variance, consistent with real mechanical gains across the fleet, because larger wheels, stronger actuators, and more drive energy do convert into more ground covered, and it would be implausible and contrary to the terramechanics literature to expect the hardware block to be inert (Ishigami et al. [\[61\]](#ref-ishigami2006); Liu et al. [\[77\]](#ref-liu2024)). Adding the autonomy block is then expected to absorb a larger incremental share and to leave the hardware coefficients attenuated, which is the signature pattern that part of what the hardware block captured in its absence was the correlated autonomy upgrade that travels with every new rover. The mechanism warranting this expectation is the collinearity of the two channels: because each new rover upgraded both hardware and autonomy at once, a regression that omits the autonomy block forces the hardware block to carry the autonomy block's explanatory weight, and restoring the autonomy block redistributes that weight to where the mechanism says it belongs. One limit deserves plain statement: because the two channels are collinear by construction at the between-rover level, this decomposition cannot by itself prove the redistribution is causal rather than an artifact of which block the analyst privileges, and for exactly this reason the within-rover contrast, not the nested decomposition, carries the primary identification. The nested decomposition is corroborating evidence whose confidence is moderate, not high, and the design is explicit that it leans on the hardware controls to separate the channels and that the separation is incomplete. Confidence would rise if the TechPort generation labels confirm that G1, G2, and G3 are genuinely distinct technology generations rather than relabelings, and would fall if the hardware and autonomy blocks prove so collinear that neither can be estimated with any precision once the other is included.

### 6.4.4 Expected terrain-interaction pattern

Illustrative, not estimated. The autonomy effect is expected to be largest in the rougher and more hazard-dense terrain classes and smallest on benign flats. The mechanism warranting this expectation is specific and is the heart of the causal story: onboard hazard detection reduces the blind-driving penalty, and that penalty is largest exactly where hazards are densest, because on hazard-dense ground a blind-commanded drive must be short and conservative to avoid stepping into an unseen hazard, whereas an autonomous drive can plan around hazards in real time and therefore cover more ground per sol; on a benign flat, by contrast, blind commanded driving is already near-optimal because there are few hazards to plan around, so the marginal value of onboard autonomy is small (Maimone, Cheng, and Matthies [\[79\]](#ref-maimone2007); Helmick et al. [\[59\]](#ref-helmick2004); Verma et al. [\[112\]](#ref-verma2025)). The operational lesson from the Curiosity traverse through Gale crater, that terrain interaction can set the productivity ceiling on the worst ground independent of the rover's intelligence, sharpens rather than contradicts this expectation: it implies that the autonomy effect should be bounded above by the terramechanical ceiling in the very roughest terrain, so the predicted interaction is increasing in roughness up to the point where wheel-soil interaction, not hazard avoidance, becomes the binding constraint (Arvidson et al. [\[18\]](#ref-arvidson2017); Golombek et al. [\[52\]](#ref-golombek2014)). This is a testable prediction with a specific shape, not a vague expectation of heterogeneity, and that specificity is what makes the third falsification check meaningful. Confidence in the interaction pattern is moderate: the mechanism is well-grounded, but the terrain-class crosswalk is a constructed artifact and the interaction is estimated on the thinnest slices of the panel, so the estimate is the most data-hungry of the three and the most exposed to the small-sample limits of the design.

### 6.4.5 What the expected signs do not claim

What the expected signs do not claim matters as much as what they do. They do not claim a specific magnitude for any coefficient, because the design stage provides no basis for a point prediction and any number offered as a magnitude would be illustration rather than forecast. They do not claim that the hardware channel is inert, because the mechanism and the terramechanics literature both imply real mechanical returns, and a finding that hardware does substantial work is fully consistent with H1 so long as the autonomy block does more. They do not claim that the autonomy effect is homogeneous, because the mechanism predicts the opposite, a heterogeneous effect concentrated in hazard-dense terrain, and the design estimates that heterogeneity rather than assuming it away with a single homogeneous coefficient. Each of these non-claims is a guard against over-reading the eventual estimates, and each is pre-specified so that the eventual reading cannot quietly expand the claim beyond what the design supports.

## 6.5 The three falsification checks, specified before estimation

The three falsification checks are specified here in operational detail, before any estimation, because naming the checks in advance is the discipline that makes the contribution falsifiable rather than merely illustrated (Angrist and Pischke [\[7\]](#ref-angrist2009)). Each check targets one component of the contribution, each has a pre-specified pass-fail criterion, and the failure of any check moves the verdict toward H0 according to the decision rule of Section 6.3.

### 6.5.1 First check: the within-rover null

The first falsification check asks whether the autonomous-drive-fraction coefficient is zero within rover. The operational criterion is that the within-rover coefficient from step five, in the specification with rover and terrain-class fixed effects, is statistically indistinguishable from zero under wild-cluster-bootstrap inference clustered at the rover level, after the measurement-error refinement of Section 6.3 has been applied. If this criterion is met, the autonomy channel has no support at the level where hardware is held constant by construction, and the contribution fails on its primary condition. This is the most direct of the three checks and is placed first in the battery because it tests the cleanest identifying variation the design offers; a contribution that cannot survive its own primary identification deserves to fall at the first check, and the design accepts that outcome in advance.

### 6.5.2 Second check: collapse of the generation block under hardware

The second falsification check asks whether the autonomy-generation coefficients lose joint significance the moment the hardware block enters the nested decomposition. The operational criterion is that the joint test of the G2-versus-G1 and G3-versus-G2 generation contrasts, significant in the specification without the hardware block, becomes insignificant under wild-cluster-bootstrap inference once the hardware block is added, and that the autonomy block's incremental explanatory share falls below the hardware block's. If this criterion is met, the apparent autonomy effect is confounded with hardware and cannot be separated from it at the between-rover level, and the contribution fails on its second condition. This check is the between-rover counterpart to the first and is acknowledged to be the weaker of the two on identification grounds, because the between-rover contrast leans on the hardware controls to separate collinear channels; its failure is therefore read in conjunction with the first check rather than in isolation, and a contribution that passes the first check but fails the second is recorded as supported at the within-rover level but unconfirmed at the between-rover level, an honest intermediate verdict rather than a forced binary.

### 6.5.3 Third check: a flat terrain-interaction profile
The third falsification check asks whether the terrain-interaction pattern is flat. The operational criterion is that the interaction of the autonomous fraction with terrain class, estimated in step five, shows no increase in the autonomy effect across terrain classes ordered from benign to hazard-dense, so that the joint test of the interaction terms is indistinguishable from zero. If this criterion is met, the proposed mechanism, that hazard avoidance reduces the blind-driving penalty most where hazards are densest, is unsupported even if a raw average autonomy effect exists, and the contribution fails on its third condition. This check distinguishes a mechanistic claim from a bare correlation. A positive average autonomy effect with a flat terrain profile would be consistent with the autonomous fraction proxying for some terrain-independent advantage of later missions, such as ground-team operational learning, rather than for the hazard-avoidance mechanism the dissertation proposes. The flat-profile check guards specifically against the operational-learning rival explanation, and its inclusion is what allows the design to claim a mechanism rather than an association.

### 6.5.4 What survives the falsification battery in each branch

The dissertation states in advance what it would conclude under each branch of the falsification battery, because a pre-registered plan should commit not only to the pass-fail criteria but to the interpretation of every reachable outcome. If all three checks pass, H1 is sustained on all three conditions and the verdict is that the autonomy channel dominates, reported with its few-cluster-aware uncertainty and its acknowledged residual-terrain-selection caveat. If the first check passes but the second fails, the verdict is that the software channel does real work within the machine but that the between-rover generation contrast cannot separate it from hardware, an intermediate finding that still supports the contribution's primary claim while conceding that the between-rover evidence is inconclusive. If the first check passes but the third fails, the verdict is that an autonomy effect exists on average but that its proposed hazard-avoidance mechanism is unsupported, which would force the dissertation to entertain the operational-learning rival as a live alternative to the mechanism and to downgrade the mechanistic claim to an associational one. If the first check fails, the contribution falls regardless of the other two, because the primary identification is the within-rover contrast and a null there cannot be rescued by between-rover evidence exposed to the very collinearity the within-rover design was built to escape. Committing to these branch interpretations in advance prevents the post-estimation rationalization that a falsified design otherwise invites, and it is the final piece of the falsifiability discipline the chapter enforces (Angrist and Pischke [\[7\]](#ref-angrist2009)).

## 6.6 Power, the minimum detectable effect, and the limits of a small panel

Before specifying the simulation and the result scaffolding, the analysis plan must confront honestly what a three-rover panel can and cannot detect, because a design that cannot in principle reject the null at a meaningful effect size is not a falsifiable design but an underpowered one masquerading as one. The current state is that the design has fixed an estimator and a decision rule; the desired state is a stated minimum detectable effect for the primary within-rover test, so that a null result can be read either as genuine evidence against H1 or as a power failure; the gap is that the panel's identifying variation is uneven across its two contrasts; and the consequence of leaving the gap unstated would be to over-claim the evidentiary weight of an eventual null.

The two contrasts have different power profiles, and the design is candid about the asymmetry. The within-rover autonomous-fraction contrast is comparatively well-powered, because its identifying variation is the large number of drive-sols within each rover, not the small number of rovers. With hundreds to thousands of drive-sols per machine across the Mars Exploration Rovers, Curiosity, and Perseverance, the within-rover regression has ample observations to estimate a slope, and its binding limit is not the count of observations but the measurement error in the reconstructed autonomous fraction and the share of within-rover variance in the fraction that survives conditioning on a priori terrain. The minimum detectable effect for the within-rover test is therefore governed by attenuation and by residual fraction variance rather than by cluster count, and the simulation of Section 6.6 quantifies the attenuation explicitly so that the realized minimum detectable effect can be reported rather than assumed. The between-rover generation contrast is badly under-powered by construction: three full panel members yield at most two generation steps, and the wild-cluster bootstrap with three clusters is conservative by design. The design states plainly that the between-rover contrast is unlikely to reject a moderate null at a conventional level and should not be expected to. This is not a flaw to be hidden but a structural fact to be reported, and it is why the decision rule makes the within-rover contrast primary and the generation contrast corroborating rather than the reverse (de Chaisemartin and D'Haultfoeuille [\[29\]](#ref-chaisemartin2020); Goodman-Bacon [\[54\]](#ref-goodmanbacon2021)).

The honest consequence for reading a null is therefore branch-specific. A null on the well-powered within-rover contrast, after the measurement-error refinement, is meaningful evidence against H1, because the test had the observations to detect a real effect and did not. A null on the under-powered between-rover generation contrast is weak evidence about anything, because the test lacked the cluster count to reject a moderate effect in the first place, and the design refuses to read such a null as support for H0. Stating this asymmetry in advance allows the eventual verdict to weight the two nulls correctly rather than treating them as equivalent, and it is a further instance of the design calibrating its epistemic claims to the evidence the panel can actually supply. The pre-registration commitment extends to the power analysis itself: the minimum-detectable-effect calculation for the within-rover test, conditional on the realized fraction variance and the simulated attenuation, is reported alongside the estimate, so that the reader can see the effect size below which a null is uninformative rather than having to infer it.

## 6.7 Illustrative simulation, descriptive generation profile and profile interpretation, and the unpopulated result tables

This section specifies the illustrative simulation that demonstrates the estimator's behavior under a known data-generating process, specifies the descriptive generation profile and autonomy-fraction-profile interpretations (neither of which identifies a causal effect of generation on productivity) that will be applied to the assembled panel, and presents the result tables as specified-but-unpopulated, with their structure fixed and their cells empty by design. Nothing in this section is an executed result; the simulation is a design artifact that exercises the estimator on synthetic data, and the empirical tables are scaffolding for results not yet obtained.

### 6.7.1 Purpose and design of the illustrative simulation

The illustrative simulation serves one purpose: to demonstrate, before any real estimation, that the specified estimator recovers a known autonomy effect under a known data-generating process and correctly returns a null when the data-generating process contains no autonomy effect. This is a property of the estimator and the inference procedure, not a forecast of the empirical result, and the simulation is reported as a methods check rather than as evidence about the Mars fleet. The synthetic data are generated to mimic the panel structure of the real design: three rovers, an unbalanced set of drive-sols per rover, a rover-level hardware block correlated with a rover-level autonomy generation, a within-rover autonomous fraction correlated with an a priori terrain class, and a terrain-class fixed effect. Two data-generating regimes are simulated. In the first regime, the synthetic autonomous fraction has a true positive effect on synthetic meters per sol that is concentrated in the rough terrain classes, matching the mechanism; the simulation verifies that the within-rover specification recovers a positive autonomous-fraction coefficient, that the nested decomposition assigns the larger incremental share to the autonomy block, and that the terrain-interaction estimate recovers the increasing profile. In the second regime, the synthetic autonomous fraction has no true effect and all of the synthetic productivity variation is generated by the hardware block; the simulation verifies that the within-rover coefficient is correctly null, that the autonomy block adds no incremental share, and that the falsification checks fire as designed. The simulation also injects classical measurement error into the synthetic autonomous fraction at several magnitudes, to quantify the attenuation the real within-rover coefficient will suffer and to calibrate the measurement-error refinement of the decision rule against a known truth.

The wild-cluster bootstrap is exercised within the simulation as well, by generating many synthetic panels under the no-effect regime and confirming that the bootstrap rejection rate at the nominal level is close to the nominal level with only three clusters, where conventional cluster-robust inference would over-reject. This is the simulation's most important methodological yield, because it converts the abstract claim that few-cluster inference requires the bootstrap into a demonstrated property of the specific estimator on the specific panel geometry, consistent with the few-cluster and heterogeneous-effects literature that motivates the choice (de Chaisemartin and D'Haultfoeuille [\[29\]](#ref-chaisemartin2020); Goodman-Bacon [\[54\]](#ref-goodmanbacon2021); Callaway and Sant'Anna [\[23\]](#ref-callawaysantanna2021)). The simulation parameters, the regimes, and the injected measurement-error magnitudes are all pre-specified, so that the simulation cannot be tuned after the fact to flatter the estimator.

### 6.7.2 The autonomous-fraction profile interpretation

The primary empirical object the design will interpret is the within-rover profile of productivity against the autonomous-drive fraction, estimated separately for each rover and each terrain class. The profile interpretation reads the slope of productivity in the autonomous fraction as the within-machine return to letting the software do more of the driving, and it reads the difference in that slope across terrain classes as the terrain-interaction pattern of Section 6.4.4. Presenting the result as a profile rather than as a single coefficient is deliberate, because the profile makes visible whether the return to autonomy is roughly linear in the fraction, as a simple model would predict, or whether it saturates, which would indicate that beyond some fraction the rover is already handling the hazards that matter and additional autonomy adds little. The profile is the natural visual companion to the within-rover coefficient and is the form in which the primary result will be communicated, with the rover-specific profiles overlaid so that the consistency of the within-machine return across the three machines can be seen directly.

### 6.7.3 The descriptive generation profile

The secondary empirical object is a descriptive generation profile that arranges the rovers along the staggered introduction of the three autonomy generations and reports the observed step changes in productivity at each generation boundary, conditional on terrain and hardware. A prefatory statement is required on what this object is and is not. No estimator in this design identifies the causal effect of rover generation on drive productivity. Every new rover upgraded its mechanical platform and its autonomy software simultaneously, so the generation boundary is confounded with the hardware transition by construction. A coefficient on a generation indicator in the between-rover specification absorbs the joint effect of the platform and the autonomy upgrade and cannot attribute it to either channel alone; even the robust estimators from the staggered-adoption literature (de Chaisemartin and D'Haultfoeuille [\[29\]](#ref-chaisemartin2020); Callaway and Sant'Anna [\[23\]](#ref-callawaysantanna2021); Goodman-Bacon [\[53\]](#ref-goodmanbacon2018)) address the heterogeneous-weighting problem in panel designs but do not resolve the hardware-autonomy collinearity that is specific to this fleet. The generation profile is therefore reported as a descriptive and associational object throughout: it describes how average productivity co-varied with the generation transition, conditional on the terrain and hardware controls, without claiming that the generation step caused the associated productivity change.

The value of the descriptive profile is contextual. It places the three generations on a common productivity axis so that the reader can see whether the step changes in observed productivity align with the generation boundaries or are better explained by the hardware covariates included in that specification. Because the two types of boundaries coincide in this fleet, the profile cannot resolve them; that is precisely the point it illustrates, and it is precisely why the within-rover autonomous-fraction contrast, where hardware is held fixed at the machine level, carries the causal weight and why the generation profile is corroborating context rather than independent identification. The profile is presented alongside the weight diagnostics for the between-rover specification so that any perverse-weight comparisons are visible rather than hidden, consistent with the few-cluster and heterogeneous-effects literature that motivates transparency about what a between-rover coefficient actually averages (de Chaisemartin and D'Haultfoeuille [\[29\]](#ref-chaisemartin2020); Goodman-Bacon [\[54\]](#ref-goodmanbacon2021)).

### 6.7.4 The result tables, specified but unpopulated by design

The result tables are specified here with their structure fixed and their cells empty, because the dissertation is a design-stage document and populating these tables with any number would violate the honesty commitment that governs the work. Each table below is named, its columns and rows are fixed, and its cells are left blank by design; the next phase of work assembles the panel and fills them.

**Table 6.1 (specified, unpopulated). Nested decomposition of productivity variance.** Rows: the three nested forms (terrain only; terrain plus hardware; terrain plus hardware plus autonomy), estimated for each of the two dependent-variable forms (meters per sol; sols per meter). Columns: incremental explanatory share of the block added at that step; the hardware coefficient block; the autonomy coefficient block; and the wild-cluster-bootstrap interval for the block added last. Every cell is empty by design at this stage.

**Table 6.2 (specified, unpopulated). Within-rover autonomous-fraction estimates.** Rows: the within-rover specification with rover and terrain-class fixed effects, estimated on the full sample and on the higher-confidence autonomous-fraction subset, for each dependent-variable form. Columns: the autonomous-fraction coefficient; its wild-cluster-bootstrap interval clustered at the rover level; and the attenuation comparison between the full sample and the high-confidence subset. Every cell is empty by design at this stage.

**Table 6.3 (specified, unpopulated). Terrain-interaction profile.** Rows: terrain classes ordered from benign to hazard-dense. Columns: the autonomous-fraction slope within each terrain class; the difference in slope from the benign baseline; and the joint test of the interaction terms. Every cell is empty by design at this stage.

**Table 6.4 (specified, unpopulated). Falsification-check verdicts.** Rows: the three falsification checks of Section 6.5. Columns: the pre-specified pass-fail criterion; the realized verdict; and the consequence for H1 under the decision rule of Section 6.3. Every cell is empty by design at this stage, and the verdict column is the one the next phase of work will populate to deliver the answer on H1 against H0.

The deliberate emptiness of these tables is not a defect of the dissertation but a feature of its honesty. A design-stage document that presented populated result tables would be claiming estimates it has not produced, and the entire credibility of a pre-registered analysis plan rests on the separation between the plan and the results. The tables are the contract: their structure is fixed now, before any estimation, so that the eventual results cannot be reshaped to fit a preferred conclusion, and their cells are filled only once the panel is assembled and the fixed procedure is run.

## 6.8 Chapter synthesis and calibrated confidence
The chapter carries the dissertation's argument through to the analysis plan. The productivity rise is genuine across the fleet, and the two channels are described together and never separated, which is exactly what the analysis plan must be built to separate (Crisp et al. [\[39\]](#ref-crisp2003); Grotzinger et al. [\[55\]](#ref-grotzinger2012); Farley et al. [\[45\]](#ref-farley2023); Gao and Chien [\[50\]](#ref-gao2021); Verma et al. [\[112\]](#ref-verma2025)). The verdict the plan delivers is decision-relevant rather than academic because sols are the binding scarcity for Mars Sample Return caching and instrument placement (Farley et al. [\[45\]](#ref-farley2023); Golombek et al. [\[52\]](#ref-golombek2014); Genova et al. [\[51\]](#ref-genova2013)). The plan reaches the mechanism rather than a confound. The within-rover autonomous-fraction contrast holds hardware fixed and isolates the software channel, and the bad-controls rule keeps realized slip on the outcome side (Angrist and Pischke [\[7\]](#ref-angrist2009); Maimone, Cheng, and Matthies [\[79\]](#ref-maimone2007); Biesiadecki and Maimone [\[20\]](#ref-biesiadecki2007); Helmick et al. [\[59\]](#ref-helmick2004)). It improves on the alternatives because the within-rover design dominates the combined-narrative description and the naive between-rover regression that cannot separate collinear channels, and the few-cluster-robust estimators dominate the naive two-way fixed-effects estimator when effects are heterogeneous (Angrist and Pischke [\[10\]](#ref-angristpischke2010); de Chaisemartin and D'Haultfoeuille [\[29\]](#ref-chaisemartin2020); Goodman-Bacon [\[54\]](#ref-goodmanbacon2021); Callaway and Sant'Anna [\[23\]](#ref-callawaysantanna2021)). The risk that remains is reportable rather than hidden: few-cluster inference via the wild-cluster bootstrap, the explicit measurement-error treatment of the reconstructed autonomous fraction, and the pre-specified falsification battery keep the residual identification risk transparent (de Chaisemartin and D'Haultfoeuille [\[29\]](#ref-chaisemartin2020); Goodman-Bacon [\[53\]](#ref-goodmanbacon2018); Imbens and Angrist [\[60\]](#ref-imbensangrist1994); Leamer [\[72\]](#ref-leamer2010)). This chapter contributes the analysis-plan portion of that argument; the conclusion draws the threads together.



Confidence in the analysis plan as a design is high. The procedure is fully specified, the decision rule is fixed and stated as a conservative conjunction, the falsification checks are named before estimation, and the inference choices follow from an evidence-rich method literature that the corpus fully supports. Confidence in the eventual verdict is, correctly, not yet warranted, because the verdict awaits the assembled panel. The design-stage honesty of the chapter consists precisely in distinguishing high confidence in the plan from no claim at all about the result. What would raise confidence in the eventual verdict is the closure of the single highest-value evidence gap, a cleaner per-drive autonomous-fraction series across all generations, which would sharpen the primary within-rover test and reduce the measurement-error attenuation that currently biases that test toward the null. What would lower it is the discovery, on panel assembly, that the autonomous-fraction reconstruction is too coarse to resolve within-rover variation, in which case the primary identification would weaken and the design would lean more heavily on the corroborating between-rover evidence whose collinearity problem it cannot fully solve. Either way, the plan specified here states in advance what it would take to move the verdict. That property distinguishes a falsifiable design from an illustrated one, and it is the contribution this chapter delivers.



# Chapter 7: Discussion

## 7.1 The chapter's answer, stated first

The verdict this design will return is decision-relevant under either outcome, and the purpose of this chapter is to make both outcomes legible in advance: to say, before a single coefficient is estimated, exactly what the world looks like if H1 holds and exactly what it looks like if H0 holds, and to interpret each of those worlds against the two anchor frameworks that organize the dissertation. That is the chapter thesis. If H1 holds, surface mobility productivity is bought more cheaply with software than with mass, the cost is paid in development and verification rather than at launch, and the capability can in principle travel to a rover already standing on Mars; the durable lever for the Mars Sample Return campaign and for the missions after it is computational. If H0 holds, the lever is mechanical, the productivity history of the fleet is a story of larger wheels and stronger actuators and more drive energy, and the design decision is frozen at the moment the spacecraft leaves Earth, because no flight-software update can grow a wheel. These two readings are not symmetric, and a central claim of this chapter is that their asymmetry is itself a reason to estimate the two channels carefully rather than to rest on the combined narrative the space-robotics literature has handed down.

The problem this chapter addresses is interpretive. Interpretation in the field today describes each Mars rover as both a better machine and a smarter one, and credits the productivity gain to the mission as a whole (Gao and Chien [\[50\]](#ref-gao2021); Farley et al. [\[45\]](#ref-farley2023)). What is missing is an interpretation that separates the two channels and tells a mission planner which lever to pull with the next increment of mass-and-power budget. The literature offers no framework for converting an estimate of relative channel contribution into a design recommendation, because it has never produced such an estimate, and even a clean estimate would be inert without a theory of why the two channels should differ in their returns and their transferability. Until that interpretive work is done, planners keep allocating budget under a narrative that cannot distinguish a recoverable error from an unrecoverable one, and the dissertation's most useful product, a decision rule for buying productivity, never gets stated. This chapter supplies that interpretation by walking both outcomes through the design-based estimation discipline and the Mokyr interpretive frame, then engaging the rivals and bounding the external validity. It does not re-derive the estimator (Chapter 5) or restate the expected signs (Chapter 6); it interprets what the eventual verdict, in either direction, would mean.

A note on register before the argument begins. Every interpretive claim in this chapter is conditioned on a result the dissertation has not yet produced. The design is at the specification stage; the panel is assembled in the next phase of work, not in this document. The chapter therefore speaks in the conditional throughout: if the within-rover autonomous-fraction coefficient survives, then the software reading follows; if it collapses, then the mechanical reading follows. Where a magnitude is named it is an illustration chosen to make a mechanism concrete, never an estimate. The confidence attached to each interpretive move is stated explicitly and is calibrated to design-stage evidence, which is to say that the confidence is in the soundness of the inference from a hypothetical result, not in the result itself.

## 7.2 Implications if H1 holds: software as the cheaper, retrofittable lever

The first interpretive claim is this: if H1 holds, the productivity history of the Mars fleet is principally a history of onboard autonomy, and the correct design response is to treat onboard computation as the primary lever for surface mobility productivity. What supports the claim is the structure of the H1 result itself. H1 is supported only if the within-rover autonomous-drive-fraction coefficient is positive and survives the inclusion of rover and terrain-class fixed effects, and if the autonomy-generation block retains joint significance and explanatory share after the hardware block enters (per the fixed decision rule of Chapter 6). The within-rover survival is the load-bearing piece, because within a single machine the wheels, the actuators, the mass, and the available drive energy are constant by construction, so the only thing that differs between a segment driven autonomously and a segment driven blind is how much of the driving the software did. A surviving within-rover coefficient therefore cannot be the hardware in disguise.

What carries that evidence to the design conclusion is the named causal mechanism the dissertation has committed to from the outset. The driver is the advance in autonomy-software generation from the Mars Exploration Rover AutoNav stack through the Mars Science Laboratory inheritance to Mars 2020 Enhanced Navigation, which processes imagery and plans a path while the rover is still in motion (Verma et al. [\[112\]](#ref-verma2025); Maimone, Cheng, and Matthies [\[79\]](#ref-maimone2007)). The mechanism is that a larger share of each drive is executed under onboard hazard detection and real-time path planning rather than blind commanded motion. The observable effect is a higher autonomous-drive fraction and more localized meters per sol, concentrated where hazards are dense. The operational consequence is fewer sols consumed per meter, which frees sols in the daily planning cycle for sampling, instrument placement, and caching. The strategic implication is that productivity bought through software costs development and verification effort but little launch mass, and can be delivered to a deployed rover through a flight-software update, as the Mars program has done across its history. This is a mechanism, not a correlation: each link names a physical or operational process, and the within-rover identification is what licenses reading the chain causally rather than as a coincidence of autonomy and hardware rising together.

That mechanism is corroborated by the engineering record of how the generations actually changed the timing and scope of computation. Biesiadecki and Maimone (2007) document the explicit tradeoff between directed (blind commanded) and autonomous driving on the Mars Exploration Rovers, which is precisely the within-rover contrast the design exploits: the same machine drove some segments under ground-specified waypoints and others under onboard hazard avoidance, and the operational difference between those modes is the software's contribution with hardware held fixed. Helmick et al. (2004) show that visual-odometry-based path following kept the rover safe in high-slip terrain where wheel odometry alone would have failed, which is the perception substrate that makes autonomous driving feasible on the ground that most throttles productivity. Verma et al. (2025) describe the qualitative step Enhanced Navigation represents: moving image processing and planning into the drive rather than between drives, so the rover no longer has to stop, think, and then move. These sources do not by themselves estimate a coefficient, and the chapter does not claim they do; their convergence is what makes the H1 mechanism mechanistically plausible rather than merely arithmetically possible, and that plausibility is what gives the causal reading its independent support.

The Mokyr reading deepens this from a statistical result into a historical pattern. Mokyr's distinction is between propositional knowledge, the understanding of why something works, and prescriptive knowledge, the technique that does the work, with the claim that progress is durable only when the prescriptive technique rests on a widening propositional base (Mokyr [\[85\]](#ref-mokyr2002)). A wheel is, in these terms, a finished prescriptive technique: its physics is fully understood, its returns are real but bounded by the same terramechanics that bounded the previous wheel, and a larger wheel is a refinement of a known kind. Autonomy software, by contrast, rests on a propositional base in perception, planning, and terrain modeling that is still widening, which is why its marginal returns can stay high and why an improvement discovered after launch can be retrofitted to a rover already on the surface. If H1 holds, the fleet's productivity history is an instance of Mokyr's pattern in which the durable gains came from the extensible, knowledge-intensive layer rather than from the bounded mechanical one (Mokyr [\[85\]](#ref-mokyr2002); Mokyr [\[86\]](#ref-mokyr2013)). Mokyr's framework also carries a warning that belongs in the H1 reading: such gains persist only when the access cost between research and deployment stays low (Mokyr [\[84\]](#ref-mokyr1998)). For the Mars fleet that access cost is the cost and risk of validating and uplinking new flight software to an operating spacecraft. The H1 design recommendation is therefore not simply to invest in autonomy but to keep the path from a better planner on the ground to a better planner on Mars cheap and low-risk, because a propositional advance frozen behind an unaffordable verification gate delivers none of its extensibility.

One qualification on the H1 interpretation is important and is stated rather than buried. The within-rover coefficient identifies the marginal productivity of shifting a segment from blind to autonomous driving on a given machine; it does not by itself prove that the **between-generation** gain was autonomy rather than the joint upgrade, because the generations changed hardware and software together. The design's two-pronged resolution (within-rover for the clean contrast, between-rover generation contrast conditioned on the hardware block for scope) is what bounds the effect, and the H1 reading is licensed at high confidence only for the within-rover claim and at moderate confidence for the broader between-generation attribution. The objection that would defeat even the within-rover claim is that the autonomous fraction is chosen by the ground team rather than assigned at random, so the segments driven autonomously may differ systematically from those driven blind in ways that terrain conditioning does not fully absorb. The design answers this with the terrain-class fixed effects and the a-priori terrain covariates and with the strict exclusion of realized slip and realized drive time from the right-hand side (Angrist and Pischke [\[7\]](#ref-angrist2009)), but residual selection on unobserved drive difficulty remains a live caveat, and the H1 interpretation inherits it.

## 7.3 Implications if H0 holds: the mechanical lever, bounded and frozen

The symmetric claim is this: if H0 holds, the productivity history of the fleet is principally mechanical, the design lever is steel and power rather than software, and the cost of every increment is paid at launch and frozen there. The basis for reading an H0 result this way is again structural. H0 stands if the autonomy-generation coefficients become jointly indistinguishable from zero once the hardware covariates enter, and if the within-rover autonomous-fraction coefficient is indistinguishable from zero after rover and terrain-class fixed effects are in place (Chapter 6 decision rule). An autonomy block that collapses the moment hardware enters means the apparent software effect was the correlated hardware upgrade wearing software's clothes, and a null within-rover coefficient means that even on a single machine, shifting a segment from blind to autonomous driving bought no measurable productivity. The reading that follows is that larger wheels, heavier and more capable actuators, and more available drive energy are what moved meters per sol, and that the autonomy advances, real as engineering achievements, were not the productivity driver the combined narrative implied.

What carries an H0 result to its design conclusion runs through a different but equally legitimate mechanism. The driver is mechanical platform growth; the mechanism is that a larger wheel and a stronger actuator convert commanded motion into actual progress more efficiently and over harsher terrain, reducing the terramechanical losses that throttle a smaller machine; the observable effect is more meters per sol that tracks hardware class rather than autonomy generation; the operational consequence is that productivity scales with the mass and power the platform carries; and the strategic implication is that the next increment of surface productivity must be bought at launch, in mass and power and therefore in launch cost and mission risk, and cannot be uploaded after landing. This is a coherent causal chain, not a residual category, and the design treats H0 as a genuine scientific possibility rather than a strawman. The terramechanics literature corroborates it: the conversion of commanded motion into actual progress is governed by wheel-soil interaction physics (Mahon et al. [\[78\]](#ref-mahon2016)), and the operational record in Gale crater shows that terrain interaction set the productivity ceiling on the worst ground regardless of how capable the autonomy stack was (Arvidson et al. [\[18\]](#ref-arvidson2017)). If H0 holds, that ceiling is mechanical, and the Curiosity megaripple experience is the type case: no amount of onboard cleverness rescues a wheel from sinking sand, and the lever that matters is the one that changes the wheel-soil contact.

The Mokyr reading of an H0 result is not that Mokyr is wrong but that this particular technology, on this particular fleet, fell on the bounded-prescriptive side of his distinction. Under H0 the mechanical refinements delivered real productivity, which is exactly what Mokyr predicts a maturing prescriptive technique can do for a time; what H0 would add is that across this fleet the autonomy layer had not yet translated its widening propositional base into deployed productivity, perhaps because the perception and planning advances were still being absorbed into operational practice rather than into measured meters per sol (Mokyr [\[85\]](#ref-mokyr2002); Mokyr [\[84\]](#ref-mokyr1998)). This is a subtle and honest reading. It does not claim that software can never be the lever; it claims that across these four rovers, on the measured productivity construct, the bounded mechanical technique was where the realized gains lived. Mokyr's own history is full of cases in which a propositional advance preceded its prescriptive payoff by decades (Mokyr [\[86\]](#ref-mokyr2013)), and an H0 verdict would locate the Mars fleet in the lag, not in a refutation of the extensibility thesis. The confidence on this reading is moderate by construction, because H0 is the rejection of the dissertation's contribution and the chapter must not over-read a null as a positive finding for the mechanical channel; a null on autonomy is consistent with a true mechanical driver and also with an autonomy effect too small or too noisily measured to detect with three full panel members.

The qualification and the objection on the H0 reading are equally explicit. The chief threat to a confident mechanical reading is measurement error in the autonomous-drive fraction. The expansion plan flags this as the single highest-value evidence gap: a clean published per-drive autonomous-fraction series across all generations is not guaranteed in the public record, and the fraction must be reconstructed from the performance reports (Verma et al. [\[112\]](#ref-verma2025); Maimone, Cheng, and Matthies [\[79\]](#ref-maimone2007); Biesiadecki and Maimone [\[20\]](#ref-biesiadecki2007)). Classical measurement error in the treatment attenuates the coefficient toward zero, which means a null within-rover coefficient could reflect a noisily measured fraction rather than a truly absent effect. The H0 interpretation therefore cannot be asserted at high confidence from a bare null; it requires the null to survive the measurement-error caveat that Chapter 5 carries and the reconstruction quality that Chapter 4 documents. This is what protects the design from declaring H0 too eagerly, and it is the reason the falsification logic is asymmetric in the right direction: rejecting the dissertation's own contribution demands a clean null, not merely an insignificant one.

## 7.4 Theoretical contribution back to each anchor framework

The dissertation does not merely borrow from its two anchors; it returns something to each, and stating what it returns is the theoretical payoff of the work regardless of the eventual verdict. The claim is that the design extends the Angrist-Pischke credibility discipline into a domain it has rarely touched and that it supplies a clean empirical test bed for the Mokyr propositional-prescriptive distinction.

To Angrist and Pischke the contribution is a demonstration that the design-based apparatus transfers from its native econometric habitat to a small-N engineering panel where the treatment and the principal confounder are upgraded together by construction (Angrist and Pischke [\[7\]](#ref-angrist2009); Angrist and Pischke [\[8\]](#ref-angrist2014)). The space-robotics literature is, in the dissertation's diagnosis, rich in description and thin in identification: it tells us repeatedly and credibly that each rover was both a better machine and a smarter one, and it never builds the comparison that would isolate either channel. The Angrist-Pischke move the dissertation makes is to find, inside an apparently hopeless collinearity, a within-unit contrast that holds the confounder fixed at the level of the individual machine, namely the autonomous-versus-blind segment comparison within a single rover. This is the credibility revolution's central instinct, that the right comparison is often hiding inside the data at a finer grain than the obvious between-unit contrast, applied to a setting where the units are spacecraft rather than people. The contribution back to the framework is the worked example: a case in which the bad-controls rule does real and non-obvious work, because the most tempting controls (realized slip, realized drive time) are exactly the post-treatment mediators that would absorb the effect, and a case in which the few-cluster problem is not a nuisance to be footnoted but a first-order constraint that forces wild-cluster bootstrap inference and an honest few-cluster caveat (Angrist and Pischke [\[7\]](#ref-angrist2009)). The dissertation thereby shows that design-based econometrics is not parochial to social science; its disciplines are exactly what a credible causal claim about a NASA technology investment requires.

To Mokyr the contribution is a rare quantitative test of a thesis usually argued from the historical record. Mokyr's propositional-prescriptive distinction and his claim that extensible, knowledge-resting techniques outpace bounded, finished ones are propositions about the long-run trajectory of technologies, and they are typically evidenced by the narrative sweep of industrial history rather than by a coefficient (Mokyr [\[85\]](#ref-mokyr2002); Mokyr [\[86\]](#ref-mokyr2013)). The Mars fleet offers something Mokyr's usual cases rarely do: two technical layers, one bounded and one extensible, advancing in the same artifact at the same time, with a public data archive that records the realized productivity of each generation. The dissertation operationalizes Mokyr's distinction as a horse race between a hardware block (the bounded prescriptive technique) and an autonomy block (the extensible technique resting on a widening propositional base), and it specifies in advance what evidence would confirm or reject the Mokyr-predicted ordering. Whatever the verdict, the design contributes a template for testing the extensibility thesis empirically rather than illustratively, and it sharpens the thesis itself by forcing a measurable construct, retrofittability, onto the otherwise abstract notion of extensibility: the autonomy layer is extensible in the specific, checkable sense that a better planner can be uploaded to a deployed rover while a better wheel cannot (Mokyr [\[84\]](#ref-mokyr1998)). The chapter claims, at moderate-to-high confidence, that this operationalization is a genuine theoretical advance for the framework, because it converts a historian's category into an estimable one without distorting it.

The two anchors do non-redundant work in the contribution exactly as they do in the design. Angrist and Pischke govern how the relative channel contribution is estimated without self-deception; Mokyr governs how the estimate is read and what it implies for the technology's future. A purely statistical result that the autonomy channel dominates would be interesting but inert without an account of why the two channels should differ in their returns and their transferability, and Mokyr supplies that account. A purely historical claim that software is the durable lever would be merely suggestive without a credible estimate that holds terrain and hardware fixed, and Angrist and Pischke supply that estimate. The dissertation's theoretical contribution is the demonstration that these two literatures, rarely in conversation, are jointly necessary to answer a concrete engineering-investment question, and that each is strengthened by the encounter.

## 7.5 Policy and mission implications for NASA, JPL, and stakeholders
The policy claim is that the verdict on H1 versus H0 is a direct input to how NASA and JPL should allocate a fixed mass-and-power budget on the next surface rover, and that the asymmetry of the policy error makes careful estimation worth its cost. This is the chapter's most operationally consequential argument, and it is argued in plain operational-economics language rather than in any architecture-traceability vocabulary, consistent with the dissertation's explicit decision to keep DoDAF/BEA framing out of a pure research-design contribution.

The mechanism by which the verdict reaches the planner is the daily planning cycle. Surface operations are organized around a cadence in which the ground team decides how far and where the rover will drive before the next communication window, and the binding scarcity in that cycle is sols (Gao and Chien [\[50\]](#ref-gao2021); Farley et al. [\[45\]](#ref-farley2023)). Mobility productivity is the currency that buys sols back: a rover that needs fewer sols per meter frees sols for the activities that constitute the science return, which in the Mars Sample Return era means sampling, instrument placement, and caching above all (Farley et al. [\[45\]](#ref-farley2023); Genova et al. [\[51\]](#ref-genova2013)). If H1 holds, the payoff of onboard autonomy is not merely the meters it adds on a given drive but the planning-cycle time it frees, because a rover trusted to handle hazards autonomously can be commanded more sparsely and less conservatively, which compounds the productivity gain through the operations process itself. This second-order effect is real and is not captured by the meters-per-sol construct the design measures, and the chapter names it as a known understatement: a confirmed autonomy effect would understate the full operational value of the autonomy channel, because the simple distance-and-sols measure does not count the ground-team planning effort consumed per meter. The design uses the simpler measure deliberately, because it is directly available from the PDS archives and is not contaminated by the analyst's model of ground-team behavior (Genova et al. [\[51\]](#ref-genova2013)), but the discussion is obligated to flag that the measured effect is a lower bound on the operational value if H1 is true.

The asymmetry of the policy error is the argument that elevates this from an interesting estimate to a decision-relevant one. Consider the two ways a planner can be wrong. If the planner wrongly concludes that productivity is mechanical when it is in fact computational, the planner over-invests in mass and power, pays the cost at launch in launch mass and mission risk, and forgoes the cheaper software lever and its retrofit option. If the planner wrongly concludes that productivity is computational when it is in fact mechanical, the planner under-builds the platform and discovers on the surface that no software update can compensate for an actuator that is too weak or a wheel that is too small. These errors are not symmetric in their recoverability. A software shortfall can sometimes be patched after landing, because flight software is uploadable; a hardware shortfall is permanent, because the wheel that left Earth is the wheel that drives on Mars. The chapter's claim, at high confidence in the logic and moderate confidence in the magnitude, is that this recoverability asymmetry is by itself a sufficient reason to estimate the two channels separately rather than to trust the combined narrative, because the combined narrative offers no way to tell the recoverable mistake from the unrecoverable one. A planner working from "the rover got better" cannot price the option value of the retrofittable lever; a planner working from a channel-separated estimate can.

The stakeholder map for this argument is concrete. The primary stakeholder is the JPL mission-formulation community designing the post-Perseverance surface architecture and the rover and fetch elements of the Mars Sample Return campaign, for whom the mass-versus-software lever is a live budget question (Farley et al. [\[45\]](#ref-farley2023)). A second stakeholder is the NASA technology-investment apparatus that funds autonomy development ahead of any specific mission, for whom the extensibility reading bears on whether autonomy research is a per-mission cost or a cross-mission asset that compounds; if H1 and the Mokyr extensibility reading hold, autonomy funding is the latter, and the access cost from research to deployed flight software is the variable to manage (Mokyr [\[84\]](#ref-mokyr1998)). A third stakeholder is the lunar surface-mobility community, including crewed-assist and uncrewed rovers, for whom the transferability of the verdict is the open question taken up in the external-validity section. The chapter does not overclaim for any of these: the design produces a Mars-fleet verdict, and the mission implications are stated as what that verdict would imply for these stakeholders, conditioned on a result not yet in hand.

A historical-operational thread is worth drawing through this, because the autonomy-versus-operations distinction predates the flight fleet the design studies. The earliest Mars surface autonomy debate is visible in the Pathfinder microrover experience, where operations and autonomy were entangled from the start and the lessons were about how much to trust the machine versus the ground (Mishkin et al. [\[83\]](#ref-mishkin2002)), and in the contemporaneous framing of autonomous rovers for Mars exploration as a way to escape the latency and bandwidth limits of ground-in-the-loop control (Washington et al. [\[116\]](#ref-washington1999)). Long-distance autonomous traverse was identified early as the specific capability that would unlock surface science at scale (Shen et al. [\[104\]](#ref-shen2003)). This lineage matters for the policy reading because it shows that the autonomy-versus-operations tradeoff the design measures is the same tradeoff the program has been negotiating since its first rover, and the dissertation's contribution is to attach an identification strategy to a question the program has answered narratively for twenty-five years. The localization machinery that makes the dependent variable measurable at all is part of that same lineage (Li et al. [\[74\]](#ref-li2002); Li et al. [\[75\]](#ref-li2011); Yang et al. [\[121\]](#ref-yang2006)), and the rapid-traverse campaigns of the current rover are the latest expression of the long-distance-autonomy ambition that the early literature named (Rankin [\[2\]](#ref-a2023)).

## 7.6 Full engagement with the rival explanations

A credible design names the comparison that would prove it wrong and the rivals that would explain the data differently, and this section takes the three principal rivals seriously rather than dismissively. The claim is that each rival is genuine, that the design has a specific answer to each, and that one of the three survives as a standing caveat the dissertation must carry honestly.

The first rival is inseparability. On this account hardware and autonomy are genuinely inseparable in this fleet because they were always upgraded together, so any attribution to one channel is an artifact of which covariate the analyst chose to privilege, and the whole enterprise of separating the channels is ill-posed. This is the strongest rival in the abstract, and the dissertation's answer is the within-rover autonomous-fraction design. Within a single machine the hardware is constant by construction, so the autonomous-versus-blind contrast varies only the software's contribution; if that contrast still shows an effect, the inseparability rival is weakened, because the effect cannot be the hardware that did not change (Biesiadecki and Maimone [\[20\]](#ref-biesiadecki2007)). What makes this answer hold is that the tradeoff between directed and autonomous driving was a real, recorded operational choice on the same rover, not a between-generation difference (Biesiadecki and Maimone [\[20\]](#ref-biesiadecki2007); Maimone, Cheng, and Matthies [\[79\]](#ref-maimone2007)). One qualification is that the within-rover design answers inseparability only for the marginal within-machine effect; it does not fully dissolve the between-generation entanglement, where hardware and autonomy did move together, and the chapter concedes that the broader generational attribution leans on the hardware controls to separate channels and is correspondingly less clean. The confidence is high that the within-rover design defeats the strong form of inseparability and moderate that it settles the generational question.

The second rival is terrain selection. On this account later rovers simply drove easier ground, so productivity rose because the terrain got friendlier, not because the rover got smarter or stronger. This rival is serious because terrain is not randomly assigned: each rover drove the terrain its landing site presented, and planners route each rover along paths chosen partly for their drivability, so observed terrain is endogenous to the rover's capability (Golombek et al. [\[52\]](#ref-golombek2014); Arvidson et al. [\[18\]](#ref-arvidson2017)). The design's answer is the terrain-class fixed effects and within-terrain-class comparison, which condition out systematic between-class differences and restrict the identifying variation to comparisons of like terrain (Golombek et al. [\[52\]](#ref-golombek2014)). The mechanism by which this works is that the productivity contrast is taken within a terrain class, so a rover that drove easier ground on average does not get credited for productivity that was really the ground's gift. The honest qualifier, stated plainly, is that terrain-class fixed effects address between-class selection but residual within-class terrain selection remains a genuine caveat: a rover may have driven the easier patches within a class, and the a-priori terrain covariates can only partly absorb that. This is the rival that survives as a standing limitation, and the chapter does not pretend otherwise; it is carried forward to the external-validity discussion and to the limitations of the conclusion as the residual that the design mitigates but does not eliminate (Golombek et al. [\[52\]](#ref-golombek2014); Arvidson et al. [\[18\]](#ref-arvidson2017)).

The third rival is ground-team operational learning. On this account the people got better, not the machine: ground teams accumulated twenty years of experience planning Mars drives, and that human learning, not onboard autonomy, drove the productivity trend. This rival is the one most easily confused with the autonomy channel, because both improve over calendar time and both reduce sols per meter. The design's answer is the specific structure of the autonomous-drive-fraction measure. Ground-team learning improves blind commanded drives too: a better-planned blind drive covers more ground than a worse-planned one, so operational learning raises productivity across both modes. The autonomy channel, by contrast, is specifically the share of the drive executed onboard, and the within-rover contrast between autonomous and blind segments differences out anything that improves both modes equally, including ground-team skill (Maimone, Cheng, and Matthies [\[79\]](#ref-maimone2007); Biesiadecki and Maimone [\[20\]](#ref-biesiadecki2007)). The mechanism is a differencing argument: if operational learning raised the productivity of blind and autonomous segments by the same amount, the autonomous-minus-blind contrast within a rover is purged of it, and a surviving contrast must be the autonomy-specific component. This holds only if operational learning improves both modes symmetrically; if ground teams learned specifically how to set up autonomous drives better over time, part of that learning would load onto the autonomy channel and inflate it. The chapter flags this asymmetric-learning possibility as a refinement of the rival rather than a refutation of the answer, and assigns moderate confidence to the claim that the differencing argument substantially, though not perfectly, separates onboard autonomy from ground-team learning.

The disposition of the three rivals is therefore not uniform, and saying so is part of the design's honesty. Inseparability is largely defeated for the marginal within-machine claim. Operational learning is largely differenced out under a symmetry assumption that is plausible but not guaranteed. Terrain selection is mitigated but survives as the standing residual caveat. A design that claimed to have killed all three would be less credible than one that names which rival it cannot fully dispatch, and the dissertation chooses the latter.

## 7.7 External-validity statement

The external-validity claim is deliberately modest: the design produces a verdict about four rovers at three landing sites on one planet, and any generalization beyond that frame is a conjecture to be tested, not a result to be asserted. Stating this bound precisely is itself a contribution, because the combined narrative the dissertation replaces tends to generalize freely from "rovers got better" to "autonomy is the future" without marking where evidence ends and extrapolation begins.

The bound follows from the structure of the panel. The fleet is small: four rovers, three full panel members with PDS traverse products, and Sojourner as a boundary case whose archive predates the traverse-product standard and which is therefore carried qualitatively rather than as a fourth panel member (the design's coverage limitation, consistent with Chapter 4). Between-rover identification rests on few clusters and leans heavily on within-rover variation, which is why the within-rover autonomous-fraction contrast is the primary test and the between-rover generation contrast the complement. With this structure, generalization to lunar rovers, to future Mars rovers with different flight processors, or to off-road terrestrial autonomy is not justified without re-estimation on those platforms. The reason for refusing to generalize is the same few-cluster, design-based discipline that governs the estimation: a claim is only as wide as the comparison that identifies it, and this comparison is Mars-fleet-specific (Angrist and Pischke [\[7\]](#ref-angrist2009)).

The Mokyr framework nonetheless supplies a structured conjecture about which way the verdict would transfer if it were tested elsewhere, and the chapter labels it a conjecture. The autonomy channel should generalize better than the hardware channel, because the propositional base underlying perception and planning is platform-general while a specific wheel or actuator is not (Mokyr [\[85\]](#ref-mokyr2002); Mokyr [\[86\]](#ref-mokyr2013)). A planner that works on Mars rests on understanding of stereo geometry, hazard modeling, and path optimization that transfers to a lunar rover or a terrestrial off-road vehicle with adaptation rather than reinvention, whereas a wheel sized for Martian regolith at Martian gravity is a finished technique tuned to its environment and does not transfer as cleanly. If that conjecture is right, then an H1 verdict on Mars would be a stronger predictor of an H1 verdict on the Moon than an H0 verdict would be, because the channel that H1 credits is the transferable one. But this is a hypothesis generated by the framework, not a finding produced by the design, and the chapter assigns it low-to-moderate confidence appropriate to an untested cross-platform extrapolation. The honest statement is that the Mars verdict bounds itself to Mars, that Mokyr gives a principled reason to expect the software reading to travel better than the hardware reading, and that the only way to know is to re-run the design on a second platform when its data archive matures. The lunar surface program and the growing literature on planetary mobility beyond the current fleet (Sopegno et al. [\[105\]](#ref-sopegno2024)) are where that replication would eventually be possible, and the conclusion names it as the natural next platform for the design rather than as a domain the present verdict already covers.

## 7.8 How this chapter advances the argument

This chapter has carried the interpretive part of the dissertation's argument, and it closes by taking stock of what the interpretation leaves unresolved. The residual risk is bounded and, more than that, transparent. The productivity rise is genuine and the channels have never been separated (Crisp et al. [\[39\]](#ref-crisp2003); Grotzinger et al. [\[55\]](#ref-grotzinger2012); Vasavada et al. [\[110\]](#ref-vasavada2014); Farley et al. [\[45\]](#ref-farley2023); Gao and Chien [\[50\]](#ref-gao2021); Verma et al. [\[112\]](#ref-verma2025)). The separation matters because sols are the binding scarcity for Mars Sample Return caching and the policy error between channels is asymmetric in recoverability (Farley et al. [\[45\]](#ref-farley2023); Genova et al. [\[51\]](#ref-genova2013); Golombek et al. [\[52\]](#ref-golombek2014)). The design reaches the mechanism because the within-rover autonomous-fraction contrast isolates the software channel with hardware held fixed and the bad-controls rule keeps realized slip on the outcome side (Angrist and Pischke [\[7\]](#ref-angrist2009); Maimone, Cheng, and Matthies [\[79\]](#ref-maimone2007); Biesiadecki and Maimone [\[20\]](#ref-biesiadecki2007); Helmick et al. [\[59\]](#ref-helmick2004)). It improves on the alternatives because the within-rover contrast dominates both the combined narrative and the naive between-rover regression that cannot separate collinear channels. What the chapter cannot fully resolve it has named plainly: residual within-class terrain selection, possible asymmetric ground-team learning, attenuation from a reconstructed autonomous-fraction measure, and the few-cluster limit on between-rover inference. None of these is hidden; each is priced into the confidence statements above; and the conclusion draws these commitments together for the dissertation as a whole.

The chapter's closing interpretive claim is that the design is decision-relevant under either verdict, which is the strongest property a falsifiable contribution can have. If H1 holds, the dissertation hands NASA and JPL a quantified case for treating onboard autonomy as the primary, retrofittable, Mokyr-extensible lever for surface productivity, with the access cost from research to flight software as the variable to manage. If H0 holds, it hands them a quantified case for treating mechanical platform capability as the binding lever that must be bought at launch, with the autonomy investment justified on grounds other than measured meters per sol. In neither case does the planner remain stranded with a narrative that cannot distinguish the recoverable mistake from the unrecoverable one. That is what the chapter set out to establish, and it is established at the design stage with the honesty the stage demands: every interpretation here is conditional on a result the next phase of work will produce, every magnitude is an illustration, and every confidence statement is calibrated to the soundness of the inference rather than to the existence of an estimate the dissertation does not yet have.



# Chapter 8: Conclusion

## 8.1 The chapter thesis

The contribution of this dissertation is a complete, falsifiable research design, not a set of executed estimates, and the value of that design does not depend on which way the eventual verdict falls. What the work delivers is a way to separate two causes of Mars surface mobility productivity that the existing literature has always reported fused together, the autonomy-software generation and the mechanical platform, and to estimate their relative contributions without fooling ourselves about which is doing the work. That separation, the named counterfactual that makes it estimable, the within-rover autonomous-fraction comparison that holds hardware fixed at the level of the individual machine, the strict bad-controls rule that keeps realized slip on the outcome side, and the asymmetric-error argument that makes the answer decision-relevant all stand whether H1 is confirmed or rejected. This chapter develops that thesis in order. It restates the contribution and isolates what survives a null result, it sets out the limitations honestly and without minimizing them, it specifies a concrete future-research program that takes the design from paper to executed estimates on the full archive, and it closes. The register remains design-stage throughout. No empirical estimate is reported anywhere in this dissertation, and every expected value that has appeared in earlier chapters is an illustration of the design rather than a finding produced from an assembled panel.

The problem this chapter addresses can be stated as a current state, a desired state, a gap, and a consequence. The current state at the close of the work is that the dissertation has specified an estimator, an identification strategy, a measurement design, an analysis plan with a fixed decision rule, and a discussion of both possible outcomes, but it has not yet run the estimation, and a reader could reasonably ask what has actually been established if no coefficient has been computed. The desired state is a clear, defensible account of what a design-stage dissertation contributes in its own right, of what would remain true even if the central hypothesis were rejected, of where the design is genuinely vulnerable, and of exactly how the next phase of work would convert the design into evidence. The gap is the one that a conclusion exists to close: the foregoing chapters built the apparatus piece by piece, and the reader needs them assembled into a single statement of what was accomplished, what was not, and what comes next. The consequence of leaving that gap open is that the work would read as an elaborate promissory note rather than as a finished and self-aware piece of scholarship, and the asymmetric policy error that motivated the whole enterprise, the error of buying mobility productivity from the wrong budget, would be left without the honest accounting of design risk that a mission planner would need before acting on the eventual estimate.

## 8.2 Restating the contribution and what stands even if H1 is not confirmed

The single falsifiable contribution is the proposition, conditional on terrain class and on the hardware covariate block, that autonomy-software generation and onboard hazard-detection capability explain the larger share of between-rover and within-rover variation in per-sol mobility productivity, against the null that productivity gains are attributable solely to mechanical platform improvement and that the autonomy-generation coefficients are jointly indistinguishable from zero once hardware covariates are included. That is H1 against H0 as stated in Chapter 1 and fixed throughout, and it is the claim the design is built to reject if the evidence runs the other way. The most important thing to say at the close is that the contribution is the design that can adjudicate this claim, not a prediction that the claim will be confirmed, and that the design has standing independent of its verdict.
Consider what survives if the design is run and H1 is rejected. The first thing that survives is the separation itself. Before this work, the space-robotics literature credited each rover's mobility record to the whole machine, describing every new rover as simultaneously a better machine and a smarter one and attributing the combined improvement to the mission as an undivided achievement (Gao and Chien [\[50\]](#ref-gao2021); Crisp et al. [\[39\]](#ref-crisp2003); Grotzinger et al. [\[55\]](#ref-grotzinger2012); Vasavada et al. [\[110\]](#ref-vasavada2014); Farley et al. [\[45\]](#ref-farley2023)). That description is accurate as far as it goes, but it is not an identification, because it never constructs a comparison in which one channel is held fixed while the other varies. Whatever the estimate turns out to be, constructing the within-rover autonomous-fraction comparison and the terrain- and hardware-conditioned nested decomposition converts an undivided narrative into two estimable quantities. This rests on the design-based discipline imported from Angrist and Pischke (2009, 2014): an empirical attribution is only as good as the counterfactual it names, and a null result from a credible design is itself a finding, because it tells a planner that the autonomy channel, net of hardware and terrain, does not carry the productivity gain. In design-based econometrics the value of an analysis lies in the validity of the comparison, not in the direction of the coefficient, so a clean null is informative in a way that a confounded positive estimate is not. One qualification must be protected: a null is informative only to the extent that the design had the power to detect an effect of operationally meaningful size, which is why Chapter 5 carried the minimum-detectable-effect discussion and the few-cluster caveat, and why a null from an underpowered design would be far weaker than a null from a powered one. A critic would press that a null could reflect measurement error in the reconstructed autonomous-drive fraction rather than the absence of an autonomy effect, and the design concedes this directly by carrying the measurement-error caveat forward as the single highest-value gap to close before execution. Confidence that the separation stands regardless of verdict is high; confidence that a null would be a clean null rather than an artifact of measurement is moderate, and it is moderate precisely because the autonomous-fraction series must be reconstructed rather than read off a clean published record.

The second thing that survives a null is the asymmetric-error argument, and it survives because it is logically prior to the estimate rather than a consequence of it. The argument, developed in Chapters 1 and 7, is that the two policy errors are not symmetric in their recoverability. A planner who wrongly concludes that productivity is mechanical when it is computational over-invests in mass and power, pays the cost at launch, and forgoes the cheaper software lever and its retrofit option. A planner who wrongly concludes that productivity is computational when it is mechanical under-builds the platform and discovers on the surface that no software update can compensate for an actuator that is too weak or a wheel that is too small. The mechanism behind the asymmetry is concrete: the Mars program has repeatedly delivered new capability to rovers already on the surface through flight-software updates, the Enhanced Navigation lineage being the most consequential recent instance (Verma et al. [\[112\]](#ref-verma2025)), whereas a wheel cannot be enlarged after launch and an actuator cannot be made stronger once the spacecraft has left Earth. The driver is the difference in the lifecycle point at which each channel's capability is fixed; the mechanism is the retrofittability of software against the immutability of mechanical structure after launch; the observable effect is that software shortfalls are sometimes patchable on the surface while hardware shortfalls are not; the operational consequence is that the cost of an estimation error differs by channel; and the strategic implication is that the two channels should be estimated separately rather than bundled, because the cost of getting the attribution wrong is borne unevenly. This argument is a reason to do the estimation well; it is not itself an estimate, and it holds whether the eventual coefficient supports H1 or H0. Confidence in the asymmetry argument is high, because it rests on the documented retrofit history of the Mars surface program and on the physics of launch rather than on any coefficient.

The third thing that survives is the measurement and identification apparatus as reusable infrastructure. The drive-sol panel constructed from the Planetary Data System traverse and localization archives, the autonomy-generation classification and autonomous-drive-fraction reconstruction from the NASA Technical Reports Server performance reports, the independent technology-readiness crosswalk from TechPort, and the terrain-class covariate construction from the terramechanics and terrain-property literature (Golombek et al. [\[52\]](#ref-golombek2014); Arvidson et al. [\[18\]](#ref-arvidson2017)) together form a measurement framework that outlives this particular hypothesis. Whatever the verdict on H1, a panel that ties each localized drive to a terrain class, an autonomy generation, an autonomous-drive fraction, and a hardware vector is an asset that future work on Mars surface mobility can reuse for questions this dissertation does not ask. Treating the apparatus as a standing contribution is justified because in empirical science the construction of a clean, public, mission-external measurement design is a contribution distinct from any single estimate produced with it, and such designs tend to be reused well beyond the question that motivated them. One qualification is that the apparatus is specified here but not yet assembled, so its standing is the standing of a fully specified protocol rather than of a delivered dataset, and that distinction is held to honestly throughout. Confidence that the specified apparatus is sound is moderate to high for the dependent variable and the hardware and terrain covariates, which rest on well-documented public archives, and moderate for the autonomous-drive fraction, which is the reconstructed quantity carrying the most measurement risk.

The fourth thing that survives is the interpretive frame, which does work regardless of the verdict. Mokyr's distinction between propositional knowledge, the understanding of why something works, and prescriptive knowledge, the technique that does the work, gives the estimate its meaning either way (Mokyr [\[85\]](#ref-mokyr2002)). If H1 is confirmed, the productivity history of the fleet is an instance of Mokyr's pattern in which the durable gains came from the extensible, knowledge-intensive layer that rests on a widening propositional base in perception and planning, rather than from the bounded refinement of a finished mechanical technique. If H1 is rejected, the same frame tells us something equally interesting, that in this particular technological episode the bounded mechanical layer carried the gain after all, which would be a genuine and reportable finding about the relative maturity of the two layers in this domain at this point in its history. Either way the frame converts a coefficient into a statement about the trajectory of the technology and about whether the lever for future productivity is extensible or bounded. An estimate without an interpretive frame is inert, and Mokyr's epistemology supplies a principled account of why the two channels should differ in their returns, an account the bare statistics cannot supply on their own. Confidence in the frame's applicability is high; confidence in which branch of it the data will select is, by construction, exactly the confidence the design exists to establish and is therefore left open.

## 8.3 Limitations, stated honestly

A design-stage dissertation owes its reader an unflinching account of where the design is weak, because the credibility of the eventual estimate depends on the threats being named in advance rather than discovered after the fact. The limitations below are not boilerplate; each one bounds the strength of the claim the executed design could support, and several were load-bearing enough to shape the estimator itself.

The first and most consequential limitation is the small cross-section of rovers. The fleet is four flight rovers, of which three are full panel members because Sojourner predates the traverse-product standard and is carried as a boundary case rather than as a fourth observation in the panel. Between-rover identification therefore rests on very few clusters, and the modern econometric literature is unambiguous that conventional cluster-robust standard errors are unreliable when the number of clusters is small (Goodman-Bacon [\[54\]](#ref-goodmanbacon2021); de Chaisemartin and D'Haultfoeuille [\[29\]](#ref-chaisemartin2020)). The mechanism by which this limitation bites is direct: with three clusters carrying the between-rover variation, the sampling distribution of the between-rover generation coefficients is poorly approximated by the usual asymptotics, so a naive standard error would overstate precision and could manufacture significance the data do not support. The design's response is to lean its primary identification on within-rover variation, where the autonomous-drive fraction differs across segments of the same machine and the cluster problem is far less severe, and to compute wild-cluster bootstrap inference for the between-rover contrasts while reporting the few-cluster caveat explicitly. This mitigates the limitation but does not dissolve it, and the honest statement is that the between-rover generation contrast is and will remain the weaker of the two identifications, reported as a complement to the within-rover result rather than as a co-equal pillar. Confidence that the within-rover identification is robust to the few-cluster problem is moderate to high; confidence in the between-rover contrast is lower and is deliberately presented as such.

The second limitation is the collinearity of autonomy generation and hardware generation, which is the central identification problem and is structural rather than incidental. Each new rover upgraded both its mechanical platform and its autonomy software at once, so at the level of the rover the two channels move together and cannot be separated by between-rover variation alone. This is not a nuisance that better data would remove; it is a feature of how the Mars program actually evolved, and it is the reason the design cannot rely on a simple between-rover regression. The within-rover autonomous-fraction contrast is the response, because within a single machine the hardware is constant by construction and only the software contribution varies across autonomous and blind segments, but the response has a cost. The autonomous fraction is chosen by the ground team rather than assigned at random, and planners drive a segment autonomously when they judge the terrain suitable and the time pressure high and drive blind when the terrain is benign or the route short and well-imaged, so the autonomous fraction is correlated with terrain and terrain affects productivity directly. The terrain-class fixed effects and the terrain covariates are therefore not optional refinements but the core of the identification, conditioning out the terrain-driven component of the autonomous-fraction choice so that the residual variation is closer to as-good-as-random with respect to productivity (Angrist and Pischke [\[7\]](#ref-angrist2009)). The honest limitation is that this conditioning can never be proven complete, that some residual correlation between the autonomous-fraction choice and unobserved terrain difficulty may remain, and that the within-rover estimate is as-good-as-random only conditional on the terrain controls being adequate. Confidence that the terrain conditioning substantially reduces the endogeneity is moderate; confidence that it removes it entirely is low, and the design does not claim it does.

The third limitation is terrain endogeneity at the mission level, which compounds the second. Each rover drove the terrain its landing site presented, so terrain and rover are entangled, and no rover was sent to a randomly chosen patch of Mars. Later rovers may simply have driven easier or harder ground than earlier ones for reasons having nothing to do with their capability, and the productivity trend could in part reflect the luck or the deliberate selection of the routes rather than any property of the machine or its software. The terrain-class fixed effects and within-terrain-class comparison address this by asking how productivity varies across drives that share a terrain class but differ in autonomy, but residual terrain selection remains a genuine caveat because the terrain classes are themselves constructed and coarse and cannot capture every dimension of drivability. The mechanism of the threat is that an omitted terrain attribute correlated with rover identity would load onto the channel coefficients and bias the attribution. The design's defense is the conditioning and the explicit acknowledgment, not a claim of immunity. Confidence that terrain selection is materially reduced by the class fixed effects is moderate; the residual risk is acknowledged rather than denied.

The fourth limitation is that meters per sol is a defensible but partial measure of productivity, and its inverse, sols per meter, inherits the same partiality. The measure counts distance and time but ignores the scientific value of where the rover went, and a rover that drove fewer meters to a higher-value target may have been more productive in the sense that matters to the mission even though it scores lower on the metric. The construct-validity threat is that the dependent variable captures locomotion efficiency rather than mission value, and the two can diverge. The design accepts this deliberately, using the simpler distance-and-sols measure because it is directly available from the PDS archives and is not contaminated by the analyst's model of scientific priority, and Chapter 7 noted that a confirmed autonomy effect would if anything understate the full operational value of the autonomy channel because it would omit the planning-cycle time that autonomy frees. The honest statement is that the dissertation answers a question about locomotion productivity, not about total mission value, and that the narrower question was chosen for measurability rather than because it is the only question worth asking. Confidence that meters per sol is a clean measure of what it measures is high; confidence that what it measures exhausts productivity in the mission-value sense is low, and the boundary is drawn explicitly.

The fifth limitation is the measurement of the treatment itself. The autonomy-generation indicator is a coarse three-level construct, G1 for the Mars Exploration Rover class with visual odometry, G2 for the Mars Science Laboratory inherited and extended stack, and G3 for the Mars 2020 Enhanced Navigation system, and the continuous autonomous-drive fraction is the finer and preferred measure (Maimone, Cheng, and Matthies [\[79\]](#ref-maimone2007); Biesiadecki and Maimone [\[20\]](#ref-biesiadecki2007); Verma et al. [\[112\]](#ref-verma2025)). The difficulty, flagged as the highest-value gap in the design, is that a clean published per-drive autonomous-fraction series across all generations is not guaranteed in the public record, so the fraction must be reconstructed from the performance reports rather than read off a ready-made table. Reconstruction introduces measurement error in exactly the variable that carries the primary identification, and classical measurement error in a right-hand-side regressor attenuates its coefficient toward zero, which would bias the test against H1 and toward a false null. The TechPort technology-readiness crosswalk is used as an external check that the generation labels are genuine distinctions rather than mission self-description, which addresses the categorical construct but not the continuous fraction. The honest statement is that the strength of the central result depends on the quality of the autonomous-fraction reconstruction, that the direction of any classical measurement-error bias is toward the null, and that closing this gap is the precondition for a strong test. Confidence in the categorical generation construct is moderate to high given the TechPort check; confidence in the reconstructed continuous fraction is moderate and is the design's most important open question.

The sixth limitation is external validity. The fleet is four rovers at three landing sites on one planet, and generalization to lunar rovers, to future Mars rovers with different flight processors, or to off-road terrestrial autonomy is not warranted without re-estimation. The Mokyr lens offers a structured conjecture that the autonomy channel should transfer better than the hardware channel, because the propositional base underlying perception and planning is platform-general while a specific wheel or actuator is not (Mokyr [\[85\]](#ref-mokyr2002)), but that conjecture is itself a hypothesis and not a result, and it would have to be tested on lunar and terrestrial platforms before it could be claimed. The honest statement is that the dissertation's findings, when produced, will speak to the Mars surface fleet and to nothing else without further work. Confidence in the internal claim about this fleet is the confidence the design is built to establish; confidence in any transfer to other platforms is low and is explicitly bracketed as conjecture.

## 8.4 A concrete future-research program

The natural next step is not a new design but the execution of this one, and the path from the present paper to executed estimates is specific enough to be laid out as a sequence of phases with named deliverables. The program below is the direct continuation of the analysis plan fixed in Chapter 6, and it is written so that a successor could pick it up without reinterpretation.

The first phase is panel assembly. It assembles the drive-sol panel from the Planetary Data System Geosciences Node traverse and localization products for the Mars Exploration Rovers, Curiosity, and Perseverance, computing for each drive-sol the localized path length and the elapsed sols so that meters per sol and its inverse are defined for every observation. The localization of surface position within orbital basemaps is the method that ties each drive to a mappable location and therefore to a terrain class, and it is the operation on which the terrain conditioning depends. The deliverable of this phase is an unbalanced panel of drive-sols nested within rovers, with the dependent variable populated and validated against published mission distance totals as an external sanity check. The mechanism by which validation works is simple: the sum of localized drive lengths over a mission should reconcile with the odometry totals reported in the mission overviews (Farley et al. [\[45\]](#ref-farley2023); Vasavada et al. [\[110\]](#ref-vasavada2014)), and a material discrepancy would signal an error in path-length extraction before any estimation is attempted. Confidence that this phase is tractable is high, because the dependent variable rests on well-documented public archives.

The second phase is treatment and covariate construction, and it is the phase that carries the most risk because it is where the autonomous-drive fraction is reconstructed. It merges the autonomy-generation classification and the autonomous-drive-fraction measures from the NTRS AutoNav and Enhanced Navigation performance reports, merges the hardware covariates and the TechPort technology-readiness records, and merges the terrain class and terrain covariates from orbital basemaps and PDS terrain characterizations. The single highest-value task in the entire program is the construction of a defensible per-drive or per-drive-sol autonomous-fraction series, because that variable carries the primary identification and because the public record does not guarantee a clean published series across all generations. The future work must therefore state explicitly how the fraction is reconstructed for each generation, must quantify the uncertainty in that reconstruction, and must carry the resulting measurement-error caveat into the estimation as a bound on the attenuation of the autonomy coefficient. A productive sub-task is to seek, through the Planetary Data System and the mission engineering teams, any drive-level telemetry distinguishing autonomously executed distance from blind commanded distance, which would replace reconstruction with measurement and would convert the design's weakest point into one of its strongest. The deliverable is a fully merged panel with the treatment, hardware, and terrain blocks populated and with a documented uncertainty model for the autonomous fraction. Confidence that this phase can be completed is moderate, and it is moderate specifically because of the autonomous-fraction reconstruction.

The third phase is estimation, and it executes the plan exactly as pre-registered so that the decision rule on H1 against H0 binds rather than being chosen after seeing the coefficients. It estimates the meters-per-sol and sols-per-meter specifications in parallel in three nested forms, terrain fixed effects only, terrain fixed effects plus hardware, and terrain fixed effects plus hardware plus autonomy, and it compares the incremental explanatory share of the autonomy block against that of the hardware block. It then estimates the within-rover autonomous-drive-fraction specification with rover fixed effects, which is the primary test of H1, and it computes wild-cluster bootstrap standard errors for the between-rover contrasts while reporting the few-cluster caveat (de Chaisemartin and D'Haultfoeuille [\[29\]](#ref-chaisemartin2020); Goodman-Bacon [\[54\]](#ref-goodmanbacon2021); Callaway and Sant'Anna [\[23\]](#ref-callawaysantanna2021)). Finally it runs the three falsification checks specified before estimation: a null within-rover autonomous-fraction coefficient rejects H1, a collapse of the autonomy-generation block when hardware enters rejects H1, and a flat terrain-interaction pattern undercuts the proposed mechanism even if a raw correlation survives. The deliverable is the populated result table that this dissertation has specified but, by design, left unpopulated, together with the verdict on the hypothesis and its honestly bounded uncertainty. Naming these checks in advance is the discipline that makes the contribution falsifiable rather than merely illustrated (Angrist and Pischke [\[7\]](#ref-angrist2009)), and carrying that discipline into execution is what distinguishes a confirmatory test from an exercise in post hoc rationalization.

Beyond executing the present design, three extensions are worth specifying because each addresses a limitation named above rather than merely adding scope. The first extension is the richer productivity construct that Chapter 7 anticipated. The present design uses distance and sols because they are directly available and uncontaminated by the analyst's model of ground-team behavior, but a confirmed autonomy effect would understate the full operational value of autonomy because a rover that can be trusted to handle hazards autonomously needs less conservative and more sparsely specified commands, which frees ground-team planning effort as well as sols (Gao and Chien [\[50\]](#ref-gao2021)). A productivity measure that counts planning effort consumed per meter alongside sols consumed per meter would capture this second-order effect, and constructing it is a concrete, if demanding, next step that would require a model of ground-team planning cost and is therefore deferred from the present measurement-clean design rather than abandoned. The mechanism it would expose is the relaxation of the daily planning-cycle scarcity, the binding constraint around which surface operations are organized, and the strategic payoff it would reveal is the part of autonomy's value that distance alone cannot see. Confidence that this construct is buildable is moderate; confidence that it would strengthen rather than weaken a confirmed autonomy effect is moderate to high on mechanistic grounds.

The second extension is replication on lunar and terrestrial platforms to test the Mokyr-structured transfer conjecture directly. The conjecture is that the autonomy channel should generalize across platforms because its propositional base in perception and planning is platform-general, while the hardware channel should not because a specific wheel or actuator is platform-specific (Mokyr [\[85\]](#ref-mokyr2002)). This is a prediction the present design cannot test, because its fleet is Mars-specific, but it is a prediction that the same estimator could test on a lunar rover program or on a terrestrial off-road autonomy fleet once comparable traverse and autonomy records exist. Specifying it here converts an external-validity caveat into a research question with a named estimator, which is the most useful thing a single-domain study can do with the limits of its own generalizability. Confidence in the conjecture is low pending data, by construction.

The third extension is to push the cross-section problem at its root by incorporating additional sources of within-fleet variation that the present design treats conservatively. Each rover's traverse spans years and many terrain regimes, and within-rover variation in autonomous fraction across that span is the design's primary identifying variation, but the panel could be deepened by exploiting flight-software updates delivered to a single rover on the surface, which change the autonomy capability while holding the hardware exactly fixed. A within-rover, pre-versus-post-update comparison around a documented autonomy upgrade would be the cleanest possible test of the software channel, because it varies the software on a fixed machine on fixed-class terrain, and the Mars program's history of on-surface software delivery makes such natural experiments at least possible to look for (Verma et al. [\[112\]](#ref-verma2025)). Whether a clean pre-post window with adequate terrain overlap exists is an empirical question for the panel-assembly phase, but it is worth specifying as the highest-value identifying variation the archive might yield. Confidence that such a window exists is low to moderate and unknown until the archive is assembled, but the payoff if it exists is high enough to justify the search.

## 8.5 Closing

This dissertation set out to answer a question that the Mars surface program has answered only as a narrative: each rover drove farther per sol than the one before it, and each was at once a better machine and a smarter one, and the combined achievement was credited to the mission as a whole. The narrative is true, and it is also not an identification. It does not tell the designer of the next rover whether to spend a fixed mass-and-power budget on heavier mechanical capability or on more onboard computation, and it cannot, because it never holds one channel fixed while the other varies. The work delivered here is the apparatus that would let that question be answered: a terrain- and hardware-conditioned panel of localized drive-sols, a within-rover autonomous-fraction comparison that holds hardware fixed at the level of the individual machine, a strict rule that keeps realized slip and realized drive time on the outcome side where they belong, few-cluster-aware inference that is honest about the smallness of the fleet, and a fixed decision rule that commits the analysis to a verdict before the coefficients are seen.

What stands at the close is not a coefficient, because none has been computed, and the dissertation has been careful to say so on every page where a number appeared. What stands is the design and the four things it carries that do not depend on its verdict: the separation of two causes the literature has always reported fused, the asymmetric-error argument that makes the separation worth doing, the reusable measurement apparatus that outlives this particular hypothesis, and the interpretive frame that converts whatever estimate emerges into a statement about whether the lever for future surface productivity is the extensible, retrofittable software layer or the bounded, launch-frozen mechanical one. The interpretation drawn from Angrist and Pischke is that a credible null is as much a finding as a credible positive, so the design is informative whichever way it resolves (Angrist and Pischke [\[7\]](#ref-angrist2009)). The interpretation drawn from Mokyr is that the answer, once produced, will say something about the maturity of the two technological layers and about which of them can still be improved after the spacecraft has left Earth (Mokyr [\[85\]](#ref-mokyr2002)).

The next phase is execution, and the path to it has been laid out as a sequence with named deliverables: assemble the panel from the public archives, reconstruct the autonomous-drive fraction with its uncertainty made explicit, estimate the pre-registered specifications, run the three falsification checks, and report the verdict on H1 against H0 with the few-cluster caveat carried through. The asymmetry that motivated the whole enterprise gives the execution its urgency. A software shortfall can sometimes be patched on the surface; a hardware shortfall cannot. That alone is reason to estimate the two channels carefully rather than to keep crediting the whole machine, and providing the design that makes careful estimation possible is the contribution this dissertation was written to deliver. The rovers themselves are the achievement of many who will never drive on Mars; the least that those who study their record can offer in return is the discipline to ask, with care and without flattery, what it was that carried them so far.
## References

[1] <span id="ref-a2021"></span> Lopez Arreguin A.. (2021). Towards in-situ characterization of regolith strength by inverse terramechanics and machine learning: A survey and applications to planetary rovers. *Planetary and Space Science*. doi: [10.1016/j.pss.2021.105271](https://doi.org/10.1016/j.pss.2021.105271).

[2] <span id="ref-a2023"></span> Rankin A.. (2023). Perseverance Rapid Traverse Campaign. *IEEE Aerospace Conference Proceedings*. doi: [10.1109/aero55745.2023.10115835](https://doi.org/10.1109/aero55745.2023.10115835).

[3] <span id="ref-abcouwer2021"></span> Neil Abcouwer; Shreyansh Daftry; Tyler Del Sesto; Olivier Toupet; Masahiro Ono; Siddarth Venkatraman; Ravi Lanka; Jialin Song. (2021). Machine Learning Based Path Planning for Improved Rover Navigation. doi: [10.1109/aero50100.2021.9438337](https://doi.org/10.1109/aero50100.2021.9438337).

[4] <span id="ref-ajr2021"></span> Lopez-Arreguin A.J.R.. (2021). Machine learning in planetary rovers: A survey of learning versus classical estimation methods in terramechanics for in situ exploration. *Journal of Terramechanics*. doi: [10.1016/j.jterra.2021.04.005](https://doi.org/10.1016/j.jterra.2021.04.005).

[5] <span id="ref-ali2006"></span> Khaled Ali; Charles A. Vanelli; Jeffrey Biesiadecki; Mark Maimone; Cheng Yang; A. Miguel San Martin; James W. Alexander. (2006). Attitude and Position Estimation on the Mars Exploration Rovers. doi: [10.1109/icsmc.2005.1571116](https://doi.org/10.1109/icsmc.2005.1571116).

[6] <span id="ref-andolfo2022"></span> Andolfo, Simone; Petricca, Flavio; Genova, Antonio. (2022). Estimation of the NASA Mars2020 Perseverance rover path through Visual Odometry. doi: [10.5194/egusphere-egu22-7695](https://doi.org/10.5194/egusphere-egu22-7695).

[7] <span id="ref-angrist2009"></span> Angrist, J. D.; Pischke, J.-S.. (2009). Mostly Harmless Econometrics: An Empiricist's Companion. *Princeton University Press*. doi: [10.1515/9781400829828](https://doi.org/10.1515/9781400829828).

[8] <span id="ref-angrist2014"></span> Angrist, J. D.; Pischke, J.-S.. (2014). Mastering 'Metrics: The Path from Cause to Effect. *Princeton University Press*. [https://press.princeton.edu/books/paperback/9780691152844/mastering-metrics](https://press.princeton.edu/books/paperback/9780691152844/mastering-metrics).

[9] <span id="ref-angristimbensrubin1996"></span> Angrist, J. D.; Imbens, G. W.; Rubin, D. B.. (1996). Identification of Causal Effects Using Instrumental Variables. *Journal of the American Statistical Association*. doi: [10.1080/01621459.1996.10476902](https://doi.org/10.1080/01621459.1996.10476902).

[10] <span id="ref-angristpischke2010"></span> Angrist, J. D.; Pischke, J.-S.. (2010). The Credibility Revolution in Empirical Economics: How Better Research Design Is Taking the Con out of Econometrics. *Journal of Economic Perspectives*. doi: [10.1257/jep.24.2.3](https://doi.org/10.1257/jep.24.2.3).

[11] <span id="ref-anon1972"></span> (1972). 132. Analysis of wheel performance (at zero slip rate) using the wheel-soil system constant. *Journal of Terramechanics*. doi: [10.1016/0022-4898(72)90071-7](https://doi.org/10.1016/0022-4898(72)90071-7).

[12] <span id="ref-anon1973"></span> (1973). 80. Analysis of wheel performance (at zero slip rate) using the wheel soil system constant. *Journal of Terramechanics*. doi: [10.1016/0022-4898(73)90118-3](https://doi.org/10.1016/0022-4898(73)90118-3).

[14] <span id="ref-anon2016b"></span> Stack, Kathryn M.; Cofield, Shannon M.; Fraeman, Abigail A.; Edwards, Christopher S.. (2016). Geologic map of the MSL Curiosity rover extended mission traverse of Aeolis Mons, Gale Crater, Mars. *Geological Society of America Abstracts with Programs*. doi: [10.1130/abs/2016am-283395](https://doi.org/10.1130/abs/2016am-283395).

[16] <span id="ref-apostolopoulos2018"></span> Apostolopoulos, Dimitrios. (2018). Results of the Inflatable Robotic Rover Testbed. *Figshare*. doi: [10.1184/r1/6560684](https://doi.org/10.1184/r1/6560684).

[17] <span id="ref-aqel2016"></span> Mohammad O. A. Aqel; Mohammad Hamiruce Marhaban; M. Iqbal Saripan; Napsiah Ismail. (2016). Review of visual odometry: types, approaches, challenges, and applications. *SpringerPlus*. doi: [10.1186/s40064-016-3573-7](https://doi.org/10.1186/s40064-016-3573-7).

[18] <span id="ref-arvidson2017"></span> Arvidson, R. E.; et al.. (2017). Mars Science Laboratory Curiosity Rover Megaripple Crossings up to Sol 710 in Gale Crater. *Journal of Field Robotics*. doi: [10.1002/rob.21647](https://doi.org/10.1002/rob.21647).

[19] <span id="ref-arvidson2017b"></span> R. E. Arvidson; Patrick DeGrosse; J. P. Grotzinger; M. Heverly; J. Shechet; Scott Moreland; M.A. Newby; N. Stein. (2017). Relating geologic units and mobility system kinematics contributing to Curiosity wheel damage at Gale Crater, Mars. *Journal of Terramechanics*. doi: [10.1016/j.jterra.2017.03.001](https://doi.org/10.1016/j.jterra.2017.03.001).

[20] <span id="ref-biesiadecki2007"></span> Jeffrey Biesiadecki; Patrick C. Leger; Mark Maimone. (2007). Tradeoffs Between Directed and Autonomous Driving on the Mars Exploration Rovers. *The International Journal of Robotics Research*. doi: [10.1177/0278364907073777](https://doi.org/10.1177/0278364907073777).

[21] <span id="ref-braunerhjelm2010"></span> Braunerhjelm, P.; Acs, Z. J.; Audretsch, D. B.; Carlsson, B.. (2010). The missing link: knowledge diffusion and entrepreneurship in endogenous growth. *Small Business Economics*. doi: [10.1007/s11187-009-9235-1](https://doi.org/10.1007/s11187-009-9235-1).

[22] <span id="ref-c2025"></span> Yeo C.. (2025). Dynamic Modeling of Wheel-Soil Lateral Interaction with Scalable Bulldozing Effects for Planetary Rover. *International Conference on Control Automation and Systems*. doi: [10.23919/iccas66577.2025.11301282](https://doi.org/10.23919/iccas66577.2025.11301282).

[23] <span id="ref-callawaysantanna2021"></span> Callaway, B.; Sant'Anna, P. H. C.. (2021). Difference-in-Differences with multiple time periods. *Journal of Econometrics*. doi: [10.1016/j.jeconom.2020.12.001](https://doi.org/10.1016/j.jeconom.2020.12.001).

[24] <span id="ref-cardone2025"></span> Cardone, Mauro. (2025). Cooperative Autonomous Navigation. *Automated and Autonomous Navigation Powered by GNSS*. doi: [10.1007/978-3-031-78753-9_11](https://doi.org/10.1007/978-3-031-78753-9_11).

[25] <span id="ref-carle2010"></span> Patrick Carle; Paul Furgale; Timothy D. Barfoot. (2010). Long-range rover localization by matching LIDAR scans to orbital elevation maps. *Journal of Field Robotics*. doi: [10.1002/rob.20336](https://doi.org/10.1002/rob.20336).

[26] <span id="ref-carsten2007"></span> Carsten, J.; Rankin, A.; Ferguson, D.; Stentz, A.. (2007). Global Path Planning on Board the Mars Exploration Rovers. *IEEE Aerospace Conference*. doi: [10.1109/aero.2007.352683](https://doi.org/10.1109/aero.2007.352683).

[27] <span id="ref-carsten2009"></span> Carsten, J.; Rankin, A.; Ferguson, D.; Stentz, A.. (2009). Global planning on the Mars Exploration Rovers: Software integration and surface testing. *Journal of Field Robotics*. doi: [10.1002/rob.20287](https://doi.org/10.1002/rob.20287).

[28] <span id="ref-chaisemartin2019"></span> de Chaisemartin, Clément; D'Haultfoeuille, Xavier. (2019). Two-way Fixed Effects Estimators with Heterogeneous Treatment Effects. doi: [10.3386/w25904](https://doi.org/10.3386/w25904).

[29] <span id="ref-chaisemartin2020"></span> de Chaisemartin, Clément; D'Haultfoeuille, Xavier. (2020). Two-Way Fixed Effects Estimators with Heterogeneous Treatment Effects. *American Economic Review*. doi: [10.1257/aer.20181169](https://doi.org/10.1257/aer.20181169).

[31] <span id="ref-charrett2009"></span> Charrett, T O H; Waugh, L; Tatam, R P. (2009). Speckle velocimetry for high accuracy odometry for a Mars exploration rover. *Measurement Science and Technology*. doi: [10.1088/0957-0233/21/2/025301](https://doi.org/10.1088/0957-0233/21/2/025301).

[32] <span id="ref-chatterjee2024"></span> Chatterjee, Pratyusha; Mitra, Debajyoti; Bhowmik, Tanima. (2024). Machine Learning-based Obstacle Avoidance Path Planning of Agricultural Rover. *2024 IEEE Calcutta Conference (CALCON)*. doi: [10.1109/calcon63337.2024.10914269](https://doi.org/10.1109/calcon63337.2024.10914269).

[33] <span id="ref-chen2025"></span> F. Chen; Natasha Jackson; Pierre Allard; Giovanni Beltrame. (2025). Path planning algorithm for a South Pole lunar rover mission. *Acta Astronautica*. doi: [10.1016/j.actaastro.2025.07.059](https://doi.org/10.1016/j.actaastro.2025.07.059).

[34] <span id="ref-cheng2006"></span> Yang Cheng; Mark Maimone; L. Matthies. (2006). Visual Odometry on the Mars Exploration Rovers. doi: [10.1109/icsmc.2005.1571261](https://doi.org/10.1109/icsmc.2005.1571261).

[35] <span id="ref-collective2020"></span> (collective). (2020). Discrete element modelling for wheel-soil interaction and the analysis of the effect of gravity. *Journal of Terramechanics*. doi: [10.1016/j.jterra.2020.06.002](https://doi.org/10.1016/j.jterra.2020.06.002).

[36] <span id="ref-collective2022"></span> (collective). (2022). Advancements in autonomous mobility of planetary wheeled mobile robots: A review. *Frontiers in Space Technologies*. doi: [10.3389/frspt.2022.1080291](https://doi.org/10.3389/frspt.2022.1080291).

[37] <span id="ref-collective2025"></span> (collective). (2025). A Comprehensive Review of Path-Planning Algorithms for Planetary Rover Exploration. *Remote Sensing*. doi: [10.3390/rs17111924](https://doi.org/10.3390/rs17111924).

[38] <span id="ref-cosenza2023"></span> Chiara Cosenza; Vincenzo Niola; Stefano Pagano; Sergio Savino. (2023). Theoretical study on a modified rocker-bogie suspension for robotic rovers. *Robotica*. doi: [10.1017/s0263574723000656](https://doi.org/10.1017/s0263574723000656).

[39] <span id="ref-crisp2003"></span> Crisp, J. A.; Adler, M.; Matijevic, J. R.; Squyres, S. W.; Arvidson, R. E.; Kass, D. M.. (2003). Mars Exploration Rover mission. *Journal of Geophysical Research: Planets*. doi: [10.1029/2002je002038](https://doi.org/10.1029/2002je002038).

[40] <span id="ref-di2011"></span> Kaichang Di; Zhaoqin Liu; Zongyu Yue. (2011). Mars Rover Localization based on Feature Matching between Ground and Orbital Imagery. *Photogrammetric Engineering & Remote Sensing*. doi: [10.14358/pers.77.8.781](https://doi.org/10.14358/pers.77.8.781).

[41] <span id="ref-di2022"></span> Di, K.; et al.. (2022). Precise pose estimation of the NASA Mars 2020 Perseverance rover through a stereo-vision-based approach. *Journal of Field Robotics*. doi: [10.1002/rob.22138](https://doi.org/10.1002/rob.22138).

[42] <span id="ref-ding2009"></span> Liang Ding; Haibo Gao; Zongquan Deng; Kazuya Yoshida; Keiji Nagatani. (2009). Slip ratio for lugged wheel of planetary rover in deformable soil: definition and estimation. doi: [10.1109/iros.2009.5354565](https://doi.org/10.1109/iros.2009.5354565).

[43] <span id="ref-ding2009b"></span> DING, Liang. (2009). Terramechanics Model for Wheel-terrain Interaction of Lunar Rover Based on Stress Distribution. *Journal of Mechanical Engineering*. doi: [10.3901/jme.2009.07.049](https://doi.org/10.3901/jme.2009.07.049).

[44] <span id="ref-ding2014"></span> Liang Ding; Zongquan Deng; Haibo Gao; Jianguo Tao; Karl Iagnemma; Guangjun Liu. (2014). Interaction Mechanics Model for Rigid Driving Wheels of Planetary Rovers Moving on Sandy Terrain with Consideration of Multiple Physical Effects. *Journal of Field Robotics*. doi: [10.1002/rob.21533](https://doi.org/10.1002/rob.21533).

[45] <span id="ref-farley2023"></span> Farley, K. A.; et al.. (2023). Overview and Results From the Mars 2020 Perseverance Rover's First Science Campaign on the Jezero Crater Floor. *Journal of Geophysical Research: Planets*. doi: [10.1029/2022je007613](https://doi.org/10.1029/2022je007613).

[46] <span id="ref-fraundorfer2012"></span> Friedrich Fraundorfer; Davide Scaramuzza. (2012). Visual Odometry: Part II: Matching, Robustness, Optimization, and Applications. *IEEE Robotics & Automation Magazine*. doi: [10.1109/mra.2012.2182810](https://doi.org/10.1109/mra.2012.2182810).

[47] <span id="ref-fujiwara2020"></span> Daisuke Fujiwara; Tetsuya Oshima; Kojiro Iizuka. (2020). Measuring the Normal Stress Distribution Acting on a Locked-Wheel of Push-Pull Locomotion Rovers via a Wheel Sensor System. *Sensors*. doi: [10.3390/s20164434](https://doi.org/10.3390/s20164434).

[48] <span id="ref-g2024b"></span> Sakayori G.. (2024). Modeling of slip rate-dependent traversability for path planning of wheeled mobile robot in sandy terrain. *Frontiers in Robotics and AI*. doi: [10.3389/frobt.2024.1320261](https://doi.org/10.3389/frobt.2024.1320261).

[49] <span id="ref-gallina2016"></span> Gallina, Alberto; Krenn, Rainer; Schäfer, Bernd. (2016). On the treatment of soft soil parameter uncertainties in planetary rover mobility simulations. *Journal of Terramechanics*. doi: [10.1016/j.jterra.2015.08.002](https://doi.org/10.1016/j.jterra.2015.08.002).

[50] <span id="ref-gao2021"></span> Gao, Y.; Chien, S.. (2021). Autonomy for Space Robots: Past, Present, and Future. *Current Robotics Reports*. doi: [10.1007/s43154-021-00057-2](https://doi.org/10.1007/s43154-021-00057-2).

[51] <span id="ref-genova2013"></span> Genova, P.; et al.. (2013). Simulations of Mars Rover Traverses. *Journal of Field Robotics*. doi: [10.1002/rob.21483](https://doi.org/10.1002/rob.21483).

[52] <span id="ref-golombek2014"></span> Golombek, M.; et al.. (2014). Terrain physical properties derived from orbital data and the first 360 sols of Mars Science Laboratory Curiosity rover observations in Gale Crater. *Journal of Geophysical Research: Planets*. doi: [10.1002/2013je004605](https://doi.org/10.1002/2013je004605).

[53] <span id="ref-goodmanbacon2018"></span> Andrew Goodman-Bacon. (2018). Difference-in-Differences with Variation in Treatment Timing. *National Bureau of Economic Research*. doi: [10.3386/w25018](https://doi.org/10.3386/w25018).

[54] <span id="ref-goodmanbacon2021"></span> Goodman-Bacon, A.. (2021). Difference-in-differences with variation in treatment timing. *Journal of Econometrics*. doi: [10.1016/j.jeconom.2021.03.014](https://doi.org/10.1016/j.jeconom.2021.03.014).

[55] <span id="ref-grotzinger2012"></span> Grotzinger, J. P.; et al.. (2012). Mars Science Laboratory Mission and Science Investigation. *Space Science Reviews*. doi: [10.1007/s11214-012-9892-2](https://doi.org/10.1007/s11214-012-9892-2).

[56] <span id="ref-h2022"></span> Yang H.. (2022). Wheel's Slip Ratio and Sinkage Estimation for Planetary Rovers Moving on Deformable Terrain. *Journal of Physics Conference Series*. doi: [10.1088/1742-6596/2203/1/012046](https://doi.org/10.1088/1742-6596/2203/1/012046).

[57] <span id="ref-haber2025"></span> Haber, James T.; Horgan, Briony; Rudolph, Amanda. (2025). Widespread Diagenesis at Unconformities in Gale Crater as Inferred From the Curiosity Rover and From Orbit. *Journal of Geophysical Research: Planets*. doi: [10.1029/2024je008304](https://doi.org/10.1029/2024je008304).

[58] <span id="ref-hayati2002"></span> S. Hayati; R. Volpe; Paul Backes; J. Balaram; Richard V. Welch; R. Ivlev; Gregory K. Tharp; Stephen Peters. (2002). The Rocky 7 rover: a Mars sciencecraft prototype. doi: [10.1109/robot.1997.619330](https://doi.org/10.1109/robot.1997.619330).

[59] <span id="ref-helmick2004"></span> Helmick, D. M.; Cheng, Y.; Clouse, D. S.; Matthies, L. H.; Roumeliotis, S. I.. (2004). Path following using visual odometry for a Mars rover in high-slip environments. *IEEE Aerospace Conference*. doi: [10.1109/aero.2004.1367679](https://doi.org/10.1109/aero.2004.1367679).

[60] <span id="ref-imbensangrist1994"></span> Imbens, G. W.; Angrist, J. D.. (1994). Identification and Estimation of Local Average Treatment Effects. *Econometrica*. doi: [10.2307/2951620](https://doi.org/10.2307/2951620).

[61] <span id="ref-ishigami2006"></span> Ishigami, G.; Miwa, A.; Nagatani, K.; Yoshida, K.. (2006). Terramechanics-Based Analysis and Traction Control of a Lunar/Planetary Rover. *Field and Service Robotics*. doi: [10.1007/10991459_22](https://doi.org/10.1007/10991459_22).

[62] <span id="ref-ishigami2006b"></span> Genya Ishigami; Keiji Nagatani; Kazuya Yoshida. (2006). Path Following Control with Slip Compensation on Loose Soil for Exploration Rover. doi: [10.1109/iros.2006.282271](https://doi.org/10.1109/iros.2006.282271).

[63] <span id="ref-ishigami2007"></span> Genya Ishigami; Akiko Miwa; Keiji Nagatani; Kazuya Yoshida. (2007). Terramechanics-based model for steering maneuver of planetary exploration rovers on loose soil. *Journal of Field Robotics*. doi: [10.1002/rob.20187](https://doi.org/10.1002/rob.20187).

[64] <span id="ref-j2020"></span> L'Haridon J.. (2020). Iron Mobility During Diagenesis at Vera Rubin Ridge, Gale Crater, Mars. *Journal of Geophysical Research Planets*. doi: [10.1029/2019je006299](https://doi.org/10.1029/2019je006299).

[65] <span id="ref-ja2022"></span> Berger J.A.. (2022). Manganese Mobility in Gale Crater, Mars: Leached Bedrock and Localized Enrichments. *Journal of Geophysical Research Planets*. doi: [10.1029/2021je007171](https://doi.org/10.1029/2021je007171).

[66] <span id="ref-jiao2010"></span> JIAO, Zhen; GAO, Haibo; DENG, Zongquan; DING, Liang. (2010). Lunar Rover Wheel-Terrain Interaction Model for Climbing-up-Slope Based on Terramechanics. *ROBOT*. doi: [10.3724/sp.j.1218.2010.00070](https://doi.org/10.3724/sp.j.1218.2010.00070).

[67] <span id="ref-johnson2008"></span> Andrew Johnson; Steven B. Goldberg; Cheng Yang; Larry Matthies. (2008). Robust and Efficient Stereo Feature Tracking for Visual Odometry. doi: [10.1109/robot.2008.4543184](https://doi.org/10.1109/robot.2008.4543184).

[68] <span id="ref-katiyar2019"></span> Katiyar, Shubhi; Dutta, Ashish. (2019). Path Planning and Obstacle Avoidance in CG Space of a 10 DOF Rover using RRT. *Proceedings of the Advances in Robotics 2019*. doi: [10.1145/3352593.3352615](https://doi.org/10.1145/3352593.3352615).

[69] <span id="ref-kruger2019"></span> Kruger, Justin; Rogg, Arno; Gonzalez, Ramon. (2019). Estimating Wheel Slip of a Planetary Exploration Rover via Unsupervised Machine Learning. *2019 IEEE Aerospace Conference*. doi: [10.1109/aero.2019.8742084](https://doi.org/10.1109/aero.2019.8742084).

[70] <span id="ref-kruuse2024"></span> Kruuse, Karin; Islam, Quazi Saimoon; Teras, Hans; Pajusalu, Mihkel. (2024). Advancing Planetary Rover Mobility: Terramechanics Wheel-Terrain Modeling in a Real-Time Simulation Framework. *IAF Space Operations Symposium*. doi: [10.52202/078367-0059](https://doi.org/10.52202/078367-0059).

[71] <span id="ref-laptre2017"></span> M. G. A. Lapôtre; B. L. Ehlmann; S. E. Minson; R. E. Arvidson; F. Ayoub; A. A. Fraeman; R. C. Ewing; N. T. Bridges. (2017). Compositional variations in sands of the Bagnold Dunes, Gale crater, Mars, from visible-shortwave infrared spectroscopy and comparison with ground truth from the Curiosity rover. *Journal of Geophysical Research Planets*. doi: [10.1002/2016je005133](https://doi.org/10.1002/2016je005133).

[72] <span id="ref-leamer2010"></span> Leamer, E. E.. (2010). Tantalus on the Road to Asymptopia. *Journal of Economic Perspectives*. doi: [10.1257/jep.24.2.31](https://doi.org/10.1257/jep.24.2.31).

[73] <span id="ref-lentaris2015"></span> George Lentaris; Ioannis Stamoulias; Dimitrios Soudris; Manolis Lourakis. (2015). HW/SW Codesign and FPGA Acceleration of Visual Odometry Algorithms for Rover Navigation on Mars. *IEEE Transactions on Circuits and Systems for Video Technology*. doi: [10.1109/tcsvt.2015.2452781](https://doi.org/10.1109/tcsvt.2015.2452781).

[74] <span id="ref-li2002"></span> Rongxing Li; Fei Ma; Fengliang Xu; Larry Matthies; Clark F. Olson; R. E. Arvidson. (2002). Localization of Mars rovers using descent and surface-based image data. *Journal of Geophysical Research Atmospheres*. doi: [10.1029/2000je001443](https://doi.org/10.1029/2000je001443).

[75] <span id="ref-li2011"></span> Rongxing Li; Shaojun He; Yunhang Chen; Min Tang; Pingbo Tang; Kaichang Di; Larry Matthies; R. E. Arvidson. (2011). MER Spirit rover localization: Comparison of ground image- and orbital image-based methods and science applications. *Journal of Geophysical Research Atmospheres*. doi: [10.1029/2010je003773](https://doi.org/10.1029/2010je003773).

[76] <span id="ref-li2015"></span> Li, Linhui; Lian, Jing; Huang, Haiyang; Wang, Hongxu; Zong, Yunpeng; Zhang, Ronghui. (2015). Traversability based obstacle avoidance path-planning and path-following control for lunar rover. *Journal of Intelligent & Fuzzy Systems*. doi: [10.3233/ifs-141331](https://doi.org/10.3233/ifs-141331).

[77] <span id="ref-liu2024"></span> Liu, J.; et al.. (2024). Current-Based Analysis and Validation of a Wheel-Soil Interaction Model for the Trafficability of a Planetary Rover. *Aerospace*. doi: [10.3390/aerospace11110892](https://doi.org/10.3390/aerospace11110892).

[78] <span id="ref-mahon2016"></span> Sadhbh Mac Mahon. (2016). Modelling and contact analysis of planetary exploration rovers. *eScholarship@McGill (McGill)*. doi: [10.82308/1711](https://doi.org/10.82308/1711).

[79] <span id="ref-maimone2007"></span> Maimone, M.; Cheng, Y.; Matthies, L.. (2007). Two years of Visual Odometry on the Mars Exploration Rovers. *Journal of Field Robotics*. doi: [10.1002/rob.20184](https://doi.org/10.1002/rob.20184).

[80] <span id="ref-maki2020"></span> Maki, J. N.; et al.. (2020). The Mars 2020 Engineering Cameras and Microphone on the Perseverance Rover: A Next-Generation Imaging System for Mars Exploration. *Space Science Reviews*. doi: [10.1007/s11214-020-00765-9](https://doi.org/10.1007/s11214-020-00765-9).

[81] <span id="ref-maurette2003"></span> M. Maurette. (2003). Mars Rover Autonomous Navigation. *Autonomous Robots*. doi: [10.1023/a:1022283719900](https://doi.org/10.1023/a:1022283719900).

[83] <span id="ref-mishkin2002"></span> Andrew Mishkin; Jack Morrison; Tam T. Nguyen; Henry W. Stone; Brian Cooper; Brian Wilcox. (2002). Experiences with operations and autonomy of the Mars Pathfinder Microrover. doi: [10.1109/aero.1998.687920](https://doi.org/10.1109/aero.1998.687920).

[84] <span id="ref-mokyr1998"></span> Mokyr, Joel. (1998). The Political Economy of Technological Change: Resistance and Innovation in Economic History. *Technological Revolutions in Europe*. doi: [10.4337/9781035335299.00013](https://doi.org/10.4337/9781035335299.00013).

[85] <span id="ref-mokyr2002"></span> Mokyr, J.. (2002). The Gifts of Athena: Historical Origins of the Knowledge Economy. *Princeton University Press*. [https://press.princeton.edu/books/paperback/9780691120133/the-gifts-of-athena](https://press.princeton.edu/books/paperback/9780691120133/the-gifts-of-athena).

[86] <span id="ref-mokyr2013"></span> Mokyr, J.. (2013). Twenty-Five Centuries of Technological Change. doi: [10.4324/9781315824284](https://doi.org/10.4324/9781315824284).

[87] <span id="ref-moore2023"></span> Moore, Joel. (2023). Aided Visual Odometry: A Sparse Bundle Adjustment Solution for Rover Navigation. doi: [10.32920/24625119.v1](https://doi.org/10.32920/24625119.v1).

[88] <span id="ref-moore2023b"></span> Moore, Joel. (2023). Aided Visual Odometry: A Sparse Bundle Adjustment Solution for Rover Navigation. doi: [10.32920/24625119](https://doi.org/10.32920/24625119).

[89] <span id="ref-motohashi2024"></span> Motohashi, Masatoshi; Kubota, Takashi. (2024). Feature Point Selection Scheme of Stereo Visual Odometry for Planetary Exploration Rover. *2024 IEEE International Conference on Advanced Intelligent Mechatronics (AIM)*. doi: [10.1109/aim55361.2024.10637076](https://doi.org/10.1109/aim55361.2024.10637076).

[90] <span id="ref-nagatani2009"></span> Keiji Nagatani; Ayako Ikeda; Keisuke Sato; Kazuya Yoshida. (2009). Accurate estimation of drawbar pull of wheeled mobile robots traversing sandy terrain using built-in force sensor array wheel. doi: [10.1109/iros.2009.5354566](https://doi.org/10.1109/iros.2009.5354566).

[92] <span id="ref-orourke2007"></span> Kevin O'Rourke; Ahmed S. Rahman; Alan M. Taylor. (2007). Trade, Knowledge, and the Industrial Revolution. *National Bureau of Economic Research*. doi: [10.3386/w13057](https://doi.org/10.3386/w13057).

[93] <span id="ref-orourke2008"></span> Kevin O'Rourke; Ahmed S. Rahman; Alan M. Taylor. (2008). Luddites and the Demographic Transition. *National Bureau of Economic Research*. doi: [10.3386/w14484](https://doi.org/10.3386/w14484).

[94] <span id="ref-pavlov2019"></span> Pavlov, Catherine; Johnson, Aaron M.. (2019). Soil Displacement Terramechanics for Wheel-Based Trenching with a Planetary Rover. *2019 International Conference on Robotics and Automation (ICRA)*. doi: [10.1109/icra.2019.8793645](https://doi.org/10.1109/icra.2019.8793645).

[95] <span id="ref-pavlov2024"></span> Pavlov, Catherine; Johnson, Aaron M.. (2024). A terramechanics model for high slip angle and skid with prediction of wheel-soil interaction geometry. *Journal of Terramechanics*. doi: [10.1016/j.jterra.2023.08.004](https://doi.org/10.1016/j.jterra.2023.08.004).

[96] <span id="ref-petersen2013"></span> Petersen, Willem; McPhee, John. (2013). A Nonlinear Volumetric Contact Model for Planetary Rover Wheel/Soil Interaction. *Volume 7A: 9th International Conference on Multibody Systems, Nonlinear Dynamics, and Control*. doi: [10.1115/detc2013-13483](https://doi.org/10.1115/detc2013-13483).

[97] <span id="ref-petersen2015"></span> Petersen, Willem; McPhee, John. (2015). Experimental Validation of a Volumetric Planetary Rover Wheel/Soil Interaction Model. *Journal of Computational and Nonlinear Dynamics*. doi: [10.1115/1.4029257](https://doi.org/10.1115/1.4029257).

[98] <span id="ref-plagarcia2024"></span> Pla-Garcia, Jorge; Rafkin, Scot; Ruíz-Pérez, María; Atreya, Sushil; Gómez, Felipe. (2024). Curiosity rover TLS-SAM measurements consistent with localized methane containment and transport by 3-D atmospheric circulation in Gale crater. doi: [10.5194/epsc2024-377](https://doi.org/10.5194/epsc2024-377).

[99] <span id="ref-ricano2025"></span> Jorge Ruben Casir Ricano; Sota Yuasa; Ryota Hino; Tomoki Koshi; Tatsuya Oyama; Kenji Nagaoka. (2025). Resilient mobility of a four-wheeled planetary rover with active suspension. *Acta Astronautica*. doi: [10.1016/j.actaastro.2025.01.030](https://doi.org/10.1016/j.actaastro.2025.01.030).

[100] <span id="ref-rusu2013"></span> RUSU, Alexandru; Moreno, Sabine; Wanatabe, Yoko; Rognant, Mathieu; Devy, Michel. (2013). Onboard Autonomous Path Planning for a Planetary Exploration Rover. *AIAA Infotech@Aerospace (I@A) Conference*. doi: [10.2514/6.2013-4825](https://doi.org/10.2514/6.2013-4825).

[101] <span id="ref-s2021"></span> Andolfo S.. (2021). Rovers localization by using 3D-to-3D and 3D-to-2D visual odometry. *2021 IEEE International Workshop on Metrology for Aerospace MetroAeroSpace 2021 Proceedings*. doi: [10.1109/metroaerospace51421.2021.9511741](https://doi.org/10.1109/metroaerospace51421.2021.9511741).

[102] <span id="ref-s2022"></span> Andolfo S.. (2022). Visual Odometry analysis of the NASA Mars 2020 Perseverance rover's images. *2022 IEEE 9th International Workshop on Metrology for Aerospace MetroAeroSpace 2022 Proceedings*. doi: [10.1109/metroaerospace54187.2022.9856188](https://doi.org/10.1109/metroaerospace54187.2022.9856188).

[103] <span id="ref-salinascamus2023"></span> Salinas-Camus, Mariana; Kulkarni, Chetan; Orchard, Marcos. (2023). Battery State-of-Health Aware Path Planning for a Mars Rover. *Annual Conference of the PHM Society*. doi: [10.36001/phmconf.2023.v15i1.3511](https://doi.org/10.36001/phmconf.2023.v15i1.3511).

[104] <span id="ref-shen2003"></span> C. N. Shen; George Nagy. (2003). Autonomous navigation to provide long-distance surface traverses for Mars Rover Sample Return Mission. doi: [10.1109/isic.1989.238673](https://doi.org/10.1109/isic.1989.238673).

[105] <span id="ref-sopegno2024"></span> Laura Sopegno; Simone Martini; Salvatore Pedone; Adriano Fagiolini; Matthew J. Rutherford; Margareta Stefanovic; Alessandro Rizzo; Patrizia Livreri. (2024). An Advanced Hexacopter for Mars Exploration: Attitude Control and Autonomous Navigation. *IEEE Transactions on Aerospace and Electronic Systems*. doi: [10.1109/taes.2024.3365667](https://doi.org/10.1109/taes.2024.3365667).

[106] <span id="ref-squyres2006"></span> S. W. Squyres; R. E. Arvidson; David Bollen; J. F. Bell; J. Brückner; Nathalie A. Cabrol; W. M. Calvin; M. H. Carr. (2006). Overview of the Opportunity Mars Exploration Rover Mission to Meridiani Planum: Eagle Crater to Purgatory Ripple. *Journal of Geophysical Research Atmospheres*. doi: [10.1029/2006je002771](https://doi.org/10.1029/2006je002771).

[107] <span id="ref-tao2022"></span> Tao, Zhuo; Zhang, Wangjun; Jia, Yang; Chen, Baichao. (2022). Path Planning Technology of Mars Rover Based on Griding of Visibility-graph Map Direction Search Method. *2022 China Automation Congress (CAC)*. doi: [10.1109/cac57257.2022.10055733](https://doi.org/10.1109/cac57257.2022.10055733).

[108] <span id="ref-toupet2020"></span> Toupet, Olivier; Del Sesto, Tyler; Ono, Masahiro; Myint, Steven; vander Hook, Joshua; McHenry, Michael. (2020). A ROS-based Simulator for Testing the Enhanced Autonomous Navigation of the Mars 2020 Rover. *2020 IEEE Aerospace Conference*. doi: [10.1109/aero47225.2020.9172345](https://doi.org/10.1109/aero47225.2020.9172345).

[110] <span id="ref-vasavada2014"></span> Vasavada, A. R.; et al.. (2014). Overview of the Mars Science Laboratory mission: Bradbury Landing to Yellowknife Bay and beyond. *Journal of Geophysical Research: Planets*. doi: [10.1002/2014je004622](https://doi.org/10.1002/2014je004622).

[111] <span id="ref-verma2023"></span> Vandi Verma; Mark Maimone; Daniel Gaines; Raymond Francis; Tara Estlin; Stephen Kuhn; Gregg Rabideau; Steve Chien. (2023). Autonomous robotics is driving Perseverance rover's progress on Mars. *Science Robotics*. doi: [10.1126/scirobotics.adi3099](https://doi.org/10.1126/scirobotics.adi3099).

[112] <span id="ref-verma2025"></span> Verma, V.; et al.. (2025). Enhanced Autonomous Navigation on the Perseverance Mars Rover. *IEEE Transactions on Field Robotics*. doi: [10.1109/tfr.2025.3636366](https://doi.org/10.1109/tfr.2025.3636366).

[113] <span id="ref-villoro1998"></span> Villoro, Luis. (1998). Personal Knowledge and Propositional Knowledge. *Belief, Personal, and Propositional Knowledge*. doi: [10.1163/9789004457898_012](https://doi.org/10.1163/9789004457898_012).

[114] <span id="ref-volpe2003"></span> R. Volpe. (2003). Mars rover navigation results using sun sensor heading determination. doi: [10.1109/iros.1999.813047](https://doi.org/10.1109/iros.1999.813047).

[115] <span id="ref-wang2025"></span> Wang, Ziyi; Yu, Di; Fathi, Mina Khalilzadeh; Pei, Chaoying. (2025). Few-Shot Learning-Enhanced Tiered Path Planning for Mars Rover Navigation. *2025 American Control Conference (ACC)*. doi: [10.23919/acc63710.2025.11107497](https://doi.org/10.23919/acc63710.2025.11107497).

[116] <span id="ref-washington1999"></span> R. Washington; Keith Golden; John Bresina; David E. Smith; C F Anderson; Trey Smith. (1999). Autonomous rovers for Mars exploration. doi: [10.1109/aero.1999.794236](https://doi.org/10.1109/aero.1999.794236).

[117] <span id="ref-wei2025"></span> Wei, Pengfei; Sun, Zezhou; Tian, He. (2025). Path planning for Mars rover with kinematic constraint based on terrain information. *2025 11th International Conference on Electrical Engineering, Control and Robotics (EECR)*. doi: [10.1109/eecr64516.2025.11077206](https://doi.org/10.1109/eecr64516.2025.11077206).

[118] <span id="ref-xiao2016"></span> XIAO, Wanshen. (2016). Design of Wheel of Manned Lunar Rover and Research on Terramechanics Model for Wheel-terrain Interaction Based on Elastic Wheel. *Journal of Mechanical Engineering*. doi: [10.3901/jme.2016.10.119](https://doi.org/10.3901/jme.2016.10.119).

[119] <span id="ref-y2011"></span> Kuwata Y.. (2011). Combined EDL-mobility planning for planetary missions. *AIAA Infotech at Aerospace Conference and Exhibit 2011*. doi: [10.2514/6.2011-1404](https://doi.org/10.2514/6.2011-1404).

[120] <span id="ref-y2023"></span> Watanabe Y.. (2023). Numerical analysis of wheel locomotion on soft soil using the extended terramechanics model based on cellular automata. *Journal of Terramechanics*. doi: [10.1016/j.jterra.2023.05.005](https://doi.org/10.1016/j.jterra.2023.05.005).

[121] <span id="ref-yang2006"></span> Cheng Yang; Mark Maimone; Larry Matthies. (2006). Visual odometry on the Mars exploration rovers - a tool to ensure accurate driving and science imaging. *IEEE Robotics & Automation Magazine*. doi: [10.1109/mra.2006.1638016](https://doi.org/10.1109/mra.2006.1638016).

[122] <span id="ref-yang2023"></span> Liwei Yang; Ping Li; Song Qian; Quan He; Jinchao Miao; Mengqi Liu; Yanpei Hu; Erexidin Memetimin. (2023). Path Planning Technique for Mobile Robots: A Review. *Machines*. doi: [10.3390/machines11100980](https://doi.org/10.3390/machines11100980).

[123] <span id="ref-yeo2025"></span> Yeo, Changmin; Görner, Martin; Seo, Younghoon; Hong, Jinsong; Oh, Sehoon. (2025). TEWD-DFO: A Soft-Soil-Aware Driving Force Observer Using Terramechanics-Enhanced Wheel Dynamics for Planetary Rover. *IECON 2025 - 51st Annual Conference of the IEEE Industrial Electronics Society*. doi: [10.1109/iecon58223.2025.11221635](https://doi.org/10.1109/iecon58223.2025.11221635).

[124] <span id="ref-yousif2015"></span> Khalid Yousif; Alireza Bab-Hadiashar; Reza Hoseinnezhad. (2015). An Overview to Visual Odometry and Visual SLAM: Applications to Mobile Robotics. *Intelligent Industrial Systems*. doi: [10.1007/s40903-015-0032-7](https://doi.org/10.1007/s40903-015-0032-7).

[125] <span id="ref-yu2021"></span> Xiaoqiang Yu; Ping Wang; Zexu Zhang. (2021). Learning-Based End-to-End Path Planning for Lunar Rovers with Safety Constraints. *Sensors*. doi: [10.3390/s21030796](https://doi.org/10.3390/s21030796).

[126] <span id="ref-z2023"></span> Wang Z.. (2023). Linear Prediction of High-Slip Sinkage for Planetary Rovers' Lugged-Wheels Based on Superposition Principle. *IEEE Robotics and Automation Letters*. doi: [10.1109/lra.2023.3236574](https://doi.org/10.1109/lra.2023.3236574).

[127] <span id="ref-zaman2022"></span> Akib Zaman; Mohammad Shahjahan Majib; Shoeb Ahmed Tanjim; Shah Md. Ahasan Siddique; Fardeen Ashraf; Shafayetul Islam; Abu Hena Md Maruf Morshed; Shadman Tajwar Shahid. (2022). Phoenix: Towards Designing and Developing a Human Assistant Rover. *IEEE Access*. doi: [10.1109/access.2022.3170908](https://doi.org/10.1109/access.2022.3170908).

[128] <span id="ref-zhang2022"></span> Tianyi Zhang; Peng Song; Yang Jia; Tian He; Junkai Sun; Chuliang Yan. (2022). Slip Estimation for Mars Rover Zhurong Based on Data Drive. *Applied Sciences*. doi: [10.3390/app12031676](https://doi.org/10.3390/app12031676).




## Appendix A. Variable and Data Dictionary

This appendix is the single authoritative operationalization table for the dissertation. Its purpose is to let a reader reconstruct every column of the analysis panel from a named archive product without consulting the author. The notation here is identical to the specification fixed in Chapter 5: for drive-sol \(i\) on rover \(r\) in terrain class \(c\), productivity is modeled as \(\text{Productivity}_{irc} = \beta_1 \, \text{AutonomyGen}_{r} + \beta_2 \, \text{Hardware}_{r} + \gamma \, \text{Terrain}_{ic} + \alpha_{r} + \delta_{c} + \epsilon_{irc}\), where \(\alpha_{r}\) is the rover fixed effect and \(\delta_{c}\) is the terrain-class fixed effect. The table below names each construct, gives its operational definition, identifies its archival source, and states its measurement scale and role in the estimating equation.

| Variable | Symbol / role | Operational definition | Source archive | Scale |
|---|---|---|---|---|
| Mobility productivity (primary) | outcome, \(\text{Productivity}\) | Localized path length driven divided by sols elapsed for the drive-sol; path length is the localized traverse path, not straight-line displacement | PDS Geosciences Node traverse and localization products | meters per sol, continuous |
| Mobility productivity (inverse) | outcome, parallel spec | Sols elapsed divided by localized path length, the planner's budgeting quantity | PDS traverse and localization products | sols per meter, continuous |
| Autonomy-software generation | treatment, \(\text{AutonomyGen}_{r}\) | Categorical generation indicator: G1 (MER-class AutoNav with visual odometry), G2 (MSL inherited and extended stack), G3 (Mars 2020 Enhanced Navigation) | NTRS performance reports [79], [20], [112]; crosswalked to TechPort | 3-level categorical |
| Autonomous-drive fraction | treatment, continuous secondary | Share of a drive's localized distance executed under onboard autonomous navigation rather than blind commanded motion | NTRS performance reports [20], [112], [111] | proportion in \([0,1]\) |
| Wheel diameter | covariate, \(\text{Hardware}_{r}\) | Nominal flight-wheel diameter from the published mission description | Mission descriptions [39], [55], [80] | continuous (meters) |
| Mass class | covariate, \(\text{Hardware}_{r}\) | Rover dry mass band from the published mission description | Mission descriptions [39], [55], [110] | ordinal class |
| Actuator class | covariate, \(\text{Hardware}_{r}\) | Drive-actuator torque/power band from the mission description | Mission descriptions [39], [80] | ordinal class |
| Available drive energy per sol | covariate, \(\text{Hardware}_{r}\) | Nominal energy available for mobility per sol from the mission power description | Mission descriptions [39], [55] | continuous (W-h) |
| Terrain class | fixed effect, \(\delta_{c}\) | A priori terrain category from orbital basemaps and PDS terrain characterizations | Orbital data and PDS terrain products [52], [18] | categorical |
| Slope | covariate, \(\text{Terrain}_{ic}\) | A priori commanded-path slope from orbital digital elevation models | Orbital DEMs [52] | continuous (degrees) |
| Physical-properties index | covariate, \(\text{Terrain}_{ic}\) | Soil bearing/strength index derived from terramechanics characterizations | Terrain-property literature [52], [61], [77] | continuous index |
| Realized slip | mediator, NOT a control | Post-drive slip from visual odometry, held strictly on the outcome side | PDS / NTRS visual-odometry products [79], [59] | proportion, excluded from RHS |
| TechPort TRL | robustness label | External technology-readiness level of each mobility technology at rover design | NASA TechPort records | ordinal (TRL 1-9) |

The single most consequential entry in this table is the row distinguishing realized slip as a mediator rather than a control. Realized slip is recorded in the same visual-odometry products that the Mars Exploration Rover navigation reports describe [79], [59], so it is tempting to place it on the right-hand side as a terrain proxy; doing so would be a bad-control error in the precise sense of Angrist and Pischke [7], because slip is produced by how the autonomy software chose to drive and would therefore absorb part of the treatment effect being estimated. The dictionary fixes this distinction at the level of the variable definition so that no downstream merge can quietly reintroduce the contaminated regressor.

A second clarification concerns the choice of two parallel outcome operationalizations rather than one. Meters per sol is the natural rate measure and is the form in which the productivity trend is usually narrated, but sols per meter is the quantity a mission planner actually budgets against a finite surface lifetime, and the two are not interchangeable under aggregation because the mean of a ratio is not the ratio of means. Estimating both in parallel is therefore not redundancy but a construct-validity check: a treatment effect that appears in the rate measure but vanishes in its inverse, or changes sign, would signal that the result is an artifact of how drive-sols are aggregated rather than a property of the underlying mobility, and the design commits in advance to reporting both. The dictionary also makes explicit that distance is the localized path length and not straight-line displacement, because a rover that drives a long arc around a hazard covers more meters than its net progress suggests, and crediting only net displacement would mechanically penalize exactly the hazard-avoidance behavior that the autonomy channel is hypothesized to improve. Using localized path length keeps the outcome measuring work done by the mobility system rather than geometric progress toward a target, which is the quantity the autonomy-versus-hardware comparison needs.

A third note records the deliberate coarseness of the hardware block. Wheel diameter, mass class, actuator class, and available drive energy are entered as the mechanical covariates because they are the platform attributes a designer trades against computation under a fixed mass-and-power budget; finer mechanical descriptors exist in the terramechanics literature [61], [77], [44], but adding them would not sharpen the autonomy-versus-hardware contrast and would worsen the collinearity between the hardware block and the generation indicator, since each new rover upgraded many mechanical attributes at once. The block is therefore kept to the decision-relevant minimum, and its limited resolution is logged as a construct-validity caveat rather than concealed.

## Appendix B. Derivations

This appendix records the two pieces of algebra that the design narrative in Chapters 5 and 6 invokes but does not write out in full: the within-rover identification of the autonomous-fraction coefficient, and the wild-cluster bootstrap procedure used for the between-rover contrasts.

**B.1 Within-rover identification.** Begin from the estimating equation with the rover fixed effect \(\alpha_{r}\) included. Apply the within transformation, subtracting the rover-specific mean from every term for rover \(r\). Because \(\text{AutonomyGen}_{r}\), the \(\text{Hardware}_{r}\) block, and \(\alpha_{r}\) are all constant within a rover, each of them is annihilated by the demeaning: their within-rover deviation is identically zero. What survives is the relation between the demeaned outcome and the demeaned autonomous-drive fraction, conditioned on demeaned terrain covariates and on the terrain-class fixed effect. Formally, letting a dot denote the within-rover mean, the surviving moment is:

\[
(\text{Productivity}_{irc} - \text{Productivity}_{\cdot rc}) = \theta \, (\text{AutonFrac}_{irc} - \text{AutonFrac}_{\cdot r}) + \gamma \, (\text{Terrain}_{ic} - \text{Terrain}_{\cdot r}) + (\delta_{c} - \delta_{\cdot r}) + u_{irc} \qquad\qquad (4)
\]

The coefficient \(\theta\) is identified from variation in the autonomous fraction across drives of the same machine, which is exactly the variation that holds hardware fixed by construction. This is the formal content of the claim that the within-rover contrast is the closest available approximation to comparing like with like [\[8\]](#ref-angrist2014). The identifying assumption made explicit by this derivation is that, after conditioning on a priori terrain, the residual variation in the autonomous fraction is mean-independent of the productivity disturbance \(u_{irc}\); the threat to that assumption is ground-team selection of when to drive autonomously, which the terrain conditioning is designed to absorb and which Chapter 5 carries as a measurement-error and selection caveat.

**B.2 Wild-cluster bootstrap.** With three full panel rovers the cluster count is far below the threshold at which conventional cluster-robust standard errors are trustworthy, so the between-rover generation contrast uses the wild-cluster bootstrap. The procedure is as follows. First, estimate the restricted model under the null that the autonomy-generation coefficient block is zero and retain the restricted residuals. Second, for each bootstrap replication, draw one Rademacher weight (plus or minus one with equal probability) per rover cluster and multiply every restricted residual in that cluster by its rover's weight, generating a bootstrap outcome. Third, re-estimate the unrestricted model on the bootstrap outcome and record the cluster-robust Wald statistic for the autonomy block. Fourth, repeat for a large number of replications and locate the observed Wald statistic in the bootstrap distribution to obtain a p-value. The Rademacher weighting at the rover level preserves the within-rover correlation structure while imposing the null, which is what gives the procedure its refinement over the asymptotic approximation when clusters are few. The few-cluster literature that motivates this choice is cited in the design chapters [28], [29], [54], [23], and the residual identification risk it leaves is reported transparently rather than assumed away.

The honest limit of B.2 deserves to be stated rather than buried. Imposing the null in the bootstrap data-generating process is what gives the restricted wild-cluster procedure its better size control, but with only three between-rover clusters the Rademacher draw admits a small finite number of distinct sign patterns, so the bootstrap distribution is coarse and the smallest attainable p-value is bounded away from zero. The practical consequence is that the between-rover generation contrast can be reported as suggestive but cannot, on its own, carry a strong rejection; this is precisely why the design assigns the primary test of H1 to the within-rover autonomous-fraction contrast of B.1, where the effective sample is drive-sols rather than rovers and the cluster scarcity does not bind in the same way. The two derivations are therefore complementary by construction: B.1 supplies the clean, hardware-fixed identification that does the inferential work, and B.2 supplies the transparent, deliberately conservative few-cluster check on the broader between-rover claim. Reporting both, and declining to lean on the contrast that the data cannot support, is how the dissertation keeps its residual identification risk transparent in this appendix.

**B.3 Terrain-interaction heterogeneity.** The proposed mechanism predicts that the autonomy effect is concentrated in hazard-dense terrain, where onboard hazard detection most reduces the blind-driving penalty, and is near zero on benign flats where blind commanded driving is already close to optimal. This prediction is operationalized, not assumed, by interacting the autonomous-fraction term with terrain class, so that the within-rover specification of B.1 estimates a separate slope per class rather than a single pooled coefficient. A flat interaction profile across classes would falsify the proposed mechanism even if a pooled effect survived, which is one of the three falsification checks fixed before estimation. Writing the heterogeneity into the estimating equation rather than testing it post hoc is what keeps the mechanism claim a genuine prediction rather than an after-the-fact rationalization, with the terrain-interaction pattern serving as the condition whose failure would overturn the mechanism.

## Appendix C. Instrument and Query Details

The three named archives are reached as follows. The dependent-variable products come from the Planetary Data System Geosciences Node, which hosts the localization and traverse products for the Mars Exploration Rovers, the Mars Science Laboratory rover Curiosity, and the Mars 2020 rover Perseverance; the localization method that ties surface position to orbital basemaps is documented in the localization literature [74], [75], [40], [25]. The traverse product to request per mission is the localized path geometry indexed by sol, from which localized path length and elapsed sols are computed directly. The autonomy-classification and autonomous-fraction measures come from the NASA Technical Reports Server; the operative records are the Mars Exploration Rover autonomous-navigation and directed-versus-autonomous tradeoff reports [79], [20] and the Perseverance Enhanced Navigation description, available both as the peer-reviewed article [112] and as the NTRS citation record [109]. The third archive, NASA TechPort, supplies the technology-readiness history used as the external generation check; TechPort entries are program records rather than peer-reviewed literature, so they are cited as data provenance and not represented as graded corpus references, which is why Appendix D crosswalks them to the generation construct rather than to a reference number. The validation step that underwrites the whole list is the resolution check: the assembled corpus carries 125 unique digital object identifiers and 128 resolvable links, and a sampled set of identifiers was confirmed against the Crossref agency endpoint, so each clickable doi.org link in the reference list above points to a live record.

The merge that assembles the panel is keyed deliberately to avoid silently mixing levels of analysis. The unit of analysis is the drive-sol: one rover, one sol, with one or more commanded drives aggregated to the sol, so the merge key is the rover-by-sol pair, and the localized path length and elapsed-sol fields from the PDS product are summed and counted respectively to that key before any other table is joined. The autonomy classification from NTRS attaches at the rover-generation level and the autonomous-fraction series attaches at the drive level, so the latter is aggregated up to the drive-sol by distance-weighting the per-drive fractions, which is the step that the measurement-error caveat in Chapter 5 attaches to and which Appendix B.1 treats as the within-rover identifying variation. Hardware covariates attach at the rover level and are therefore constant within the rover fixed effect by construction, which is exactly why they drop out under the within transformation of B.1 and why the between-rover specification must instead lean on the terrain-class fixed effect plus the explicit hardware block to keep them identified. Terrain class and the continuous terrain covariates attach at the drive-sol level through the a priori commanded path, never through the realized track, preserving the bad-controls boundary at the point of the join rather than only at the point of estimation. A secondary panel keyed to the individual drive, used for the robustness battery, is built from the same products at finer resolution where the archive resolves multiple drives within a sol; it is documented here so that the drive-level robustness results can be reproduced from the same access paths without a separate data request.

One provenance limit is recorded for completeness. Two vault literature services that a full systematic sweep would normally include did not respond during corpus assembly: Semantic Scholar returned an expired host certificate and the IEEE Xplore developer key was inactive, while OpenAlex, Crossref, NTRS, and Scopus did respond and supplied the corpus. This is logged so that a later replication knows which sources were and were not consulted, and the absence is a coverage caveat on the literature sweep rather than a defect in any cited entry, every one of which resolved.

## Appendix D. Supplementary Tables

The supplementary table below is the autonomy-generation crosswalk that lets a reader confirm the three-level treatment construct is not mission marketing but a genuine technological progression with an external maturity anchor. Each generation is tied to its defining capability, the flight rovers that carried it, the primary corpus evidence, and the external TechPort technology-readiness anchor.

| Generation | Defining capability | Flight rovers | Primary evidence | External TRL anchor |
|---|---|---|---|---|
| G1 | MER-class AutoNav: onboard stereo hazard assessment, visual odometry for slip-aware position, global path planning | Spirit, Opportunity | [79], [20], [59], [27], [26] | TechPort mobility-technology records, MER era |
| G2 | Inherited and extended MER stack on a larger platform; terrain interaction as the productivity ceiling | Curiosity | [55], [110], [18], [19], [52] | TechPort mobility-technology records, MSL era |
| G3 | Enhanced Navigation: image processing and path planning executed while driving, raising autonomous-drive fraction | Perseverance | [112], [111], [108], [80], [41] | TechPort mobility-technology records, Mars 2020 era |

The crosswalk does the assurance work that closes this back matter: it shows that the treatment variable rests on an external maturity record as well as on mission self-description, which is the robustness commitment the design chapter made when it promised that the generation labels would be checked against TechPort rather than taken on the missions' word. Read together with Appendix A, it establishes that the autonomy-versus-hardware comparison is built from constructs that are each independently sourced, which is the precondition for the within-rover identification in Appendix B to mean what the dissertation claims it means. Two evidence gaps acknowledged in the expansion plan are restated here for honesty. First, a clean published per-drive autonomous-fraction series across all three generations is not guaranteed in the public record, so the fraction is reconstructed from the performance reports under the measurement-error caveat carried in Chapter 5. Second, Sojourner predates the PDS traverse-product standard and is therefore treated as a qualitative boundary case rather than a fourth panel member, and no value has been fabricated to fill that gap. The confidence attached to the crosswalk itself is high for the G1 through G3 ordering, which the cited literature and the external TechPort anchor jointly support, and moderate for any future quantitative weighting of the generations, which awaits the executed estimation specified in the analysis plan.