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

**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)
**Date:** 2026-06-15

---

## Abstract

Each successive Mars surface mission has driven farther per sol than the one before it, and the four flight rovers (Sojourner, the two Mars Exploration Rovers, Curiosity, and Perseverance) differ from one another on two axes at once: their mechanical platforms grew larger and more capable, and their onboard autonomy software advanced through several distinct generations. The standard narrative attributes the growth in mobility productivity to better hardware, larger wheels, more powerful actuators, and a heavier chassis. This dissertation states and tests 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, the autonomy-generation coefficient 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 improvements. The method is a panel regression of per-sol traverse productivity across the fleet on autonomy-generation indicators, terrain covariates, and hardware covariates, with rover and terrain-class fixed effects. Data are drawn from the Planetary Data System rover traverse and localization archives for the Mars Exploration Rovers, Curiosity, and Perseverance, supplemented by NASA Technical Reports Server performance reports for AutoNav and the Enhanced Navigation system and by TechPort technology-readiness records for mobility technologies. This document is a design-stage dissertation. It specifies the estimator, the identification strategy, and the threats to validity in full, and it presents expected results as clearly labeled illustrations rather than as estimates produced from the assembled panel. The work matters because mission planners need to 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 sample-return architectures.

---

## 1. Introduction and Contribution

Mobility is the binding constraint on what a planetary rover can accomplish. A rover that cannot reach a target cannot sample it, and the number of sols consumed reaching a target is time not spent on science. Across the Mars surface program the trend in mobility productivity has been steeply upward. Sojourner, in 1997, traveled on the order of one hundred meters across its entire mission. The Mars Exploration Rovers, Spirit and Opportunity, drove tens of kilometers over many years. Curiosity has driven more than thirty kilometers, and Perseverance has set single-sol and multi-sol distance records that exceed anything earlier in the program. The question this dissertation asks is not whether productivity rose. It plainly did. The question is what caused it to rise.

Two explanations compete. The first is mechanical. Later rovers are larger, carry more capable actuators, use larger wheels, and have more available energy. On this account, the rover simply became a better machine, and a better machine covers more ground. The second explanation is computational. Each rover generation introduced new autonomy software. The Mars Exploration Rovers flew an autonomous navigation capability built around stereo hazard assessment and visual odometry (Maimone, Cheng, and Matthies 2007; Biesiadecki and Maimone 2006). Curiosity inherited and extended that stack. Perseverance introduced Enhanced Navigation, often described as the ability to think while driving, which substantially raised the fraction of a drive that the rover could plan and execute autonomously rather than under blind commanded motion (Verma et al. 2025). On this account, the rover became smarter, and a smarter rover spends fewer sols per meter because it can plan around hazards in real time rather than waiting for ground-in-the-loop commands.

These two explanations are usually presented together and rarely separated. The literature describes each new rover as both a better machine and a smarter one, and credits the combined improvement to the mission as a whole. That description is accurate but it is not an identification. It does not tell a designer of the next rover how to allocate a fixed mass and power budget between heavier mechanical capability and more onboard computation. The gap this dissertation fills is the absence of a design that holds terrain and hardware fixed and asks how much of the productivity gain is attributable to the autonomy-software generation alone.

The single falsifiable contribution is stated as a pair of hypotheses.

- **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 contribution is falsifiable in the strong sense that the design specifies in advance what evidence would reject it. If the autonomy-generation block loses significance and explanatory share when hardware covariates enter, H1 is rejected and H0 stands. The design also specifies the rival explanations and the diagnostics that distinguish them, following the practice that a credible empirical claim must name the comparison that would prove it wrong (Angrist and Pischke 2009).

Why this matters for NASA and JPL is direct. Mobility productivity is a currency that planners spend. A rover that needs fewer sols per meter frees sols for sampling, for instrument placement, and for caching, which is the central activity of the Mars Sample Return campaign. If productivity is bought with mass and power, then the design lever is mechanical and the cost is paid at launch. If productivity is bought with software, then the design lever is computational, the cost is paid in development and verification, and the capability can in principle be uploaded to a rover already on the surface. Knowing which lever dominates is a precondition for rational investment in future surface systems.

---

## 2. Background and Literature

### 2.1 The mobility systems and their autonomy

The Mars Exploration Rover mission established the baseline for autonomous surface mobility on Mars (Crisp et al. 2003). Its navigation stack combined stereo-vision hazard assessment with onboard path selection and with visual odometry for slip-aware position estimation. Maimone, Cheng, and Matthies (2007) report two years of visual odometry operations on Spirit and Opportunity and document how slip estimation changed what drives were feasible on sloped and sandy terrain. Biesiadecki and Maimone (2006) describe the rover surface-mobility software and the trade between blind commanded driving and autonomous hazard-avoidance driving, which is precisely the trade this dissertation seeks to quantify across generations. Carsten et al. (2009) document the integration and surface testing of global path planning, the Field D* planner, on the Mars Exploration Rovers, and Carsten, Rankin, and others (2007) describe the global path-planning capability onboard. Helmick et al. (2004) show path following using visual odometry in high-slip environments, the failure mode that most directly couples terrain to productivity.

Curiosity, the Mars Science Laboratory rover, inherited this lineage and extended it (Grotzinger et al. 2012; Vasavada et al. 2014). Its larger mass and wheels changed the mechanical envelope, but the autonomy stack remained recognizably descended from the Mars Exploration Rover software. Terrain interaction became a first-order operational concern: Arvidson et al. (2017) and the megaripple-crossing analysis of the Curiosity traverse through Gale crater (Arvidson et al. 2016) show how dune and ripple fields impose slip and wheel wear that throttle productivity independent of the rover's intelligence. The terrain physical properties derived in the first sols of the mission (Golombek et al. 2014) provide the kind of terrain covariate this dissertation requires.

Perseverance introduced the most consequential autonomy step. Verma et al. (2025) describe Enhanced Navigation, the system that lets the rover process imagery and plan while it is still in motion, raising the autonomous drive fraction and the achievable distance per sol. The Mars 2020 engineering camera system (Maki et al. 2020) and the stereo-vision pose-estimation pipeline (Di et al. 2022) supply the perception substrate for that capability. Perseverance's first-campaign overview (Farley et al. 2023) records traverse and sampling activity against the Jezero crater floor.

Reviews place these systems in a longer trajectory. Gao and Chien (2021) survey autonomy for space robots across past, present, and future. The survey of advancements in autonomous mobility of planetary wheeled robots (2022) and the comprehensive review of path-planning algorithms for planetary rover exploration (2025) catalog the algorithmic generations from the early reactive hazard-avoidance schemes through the global planners and on to the learning-augmented planners now in development. On the mechanical side, terramechanics research models the wheel-soil interaction that converts commanded motion into actual progress (Ishigami et al. 2006; the wheel-soil interaction validation of Liu et al. 2024; discrete-element modeling of wheel-soil interaction under varying gravity, 2020). This body of work is the source of the hardware covariates and the terrain covariates that the design must hold fixed.

It is worth being precise about what each generation changed, because the design depends on classifying the autonomy generation correctly. The Mars Exploration Rover generation introduced two capabilities that mattered for productivity. The first was onboard stereo hazard assessment, which let the rover build a local terrain map and reject unsafe steps without a ground decision. The second was visual odometry, which let the rover measure its own slip by tracking surface features between images and thereby drive safely on slopes and in sand where wheel odometry alone would have been dangerously wrong (Maimone, Cheng, and Matthies 2007; Helmick et al. 2004). The addition of a global path planner, integrated and tested on the surface, extended the rover's 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 (Carsten et al. 2009). The Mars Science Laboratory generation carried this stack forward onto a much larger platform and into the harsher wheel-wear environment of Gale crater, where the operational lesson was that terrain interaction, not raw computing, set the productivity ceiling on the worst ground (Arvidson et al. 2016, 2017). The Mars 2020 generation changed the timing of computation itself: Enhanced Navigation moved image processing and path planning into the drive rather than between drives, so that the rover no longer had to stop, think, and then move, which is the change that most plausibly raised meters per sol on hazard-dense ground (Verma et al. 2025). This progression is the spine of the autonomy-generation variable, and the TechPort technology-readiness records provide an external check that these are genuinely distinct generations rather than relabelings of a single capability.

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

The space-robotics literature is 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 that productivity rose. It does not isolate the causal contribution of any one factor because it does not construct a comparison in which the other factors are held fixed. This is exactly the situation that design-based econometrics was built to address (Angrist and Pischke 2009, 2014).

The Angrist-Pischke apparatus supplies three disciplines that this dissertation adopts in plain language. First, every causal claim must name the counterfactual and the comparison that identifies it. Here the comparison is variation in per-sol productivity across drives that share a terrain class but differ in autonomy generation, with rover identity itself absorbed by fixed effects where the variation is within-rover. Second, the design must guard against bad controls, that is, against conditioning on variables that are themselves consequences of the treatment (Angrist and Pischke 2009). Realized slip, for example, is partly a function of how the autonomy software chose to drive; conditioning on realized slip would absorb part of the effect of interest and bias the estimate toward the null. The design therefore treats commanded-drive characteristics and a priori terrain class as covariates and treats realized slip as an outcome-side mediator, not a control. Third, inference must be honest about the panel structure. With repeated drives nested within a small number of rovers, the effective number of independent clusters is small, and conventional standard errors understate uncertainty. The recent literature on fixed-effects estimation with few clusters and with heterogeneous and dynamic effects (de Chaisemartin and D'Haultfoeuille 2020; Goodman-Bacon 2021) bears directly on how the autonomy-generation coefficients should be estimated and how their standard errors should be computed.

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

Mokyr's economic history of technology supplies the second lens (Mokyr 2002). His central distinction is between propositional knowledge, the understanding of why something works, and prescriptive knowledge, the technique that does the work. Progress is sustained, Mokyr argues, only when the prescriptive technique rests on a widening propositional base, because techniques discovered by trial without underlying theory tend to stagnate, while techniques that rest on deep understanding are extensible and self-correcting.

This frames the autonomy-versus-hardware question historically rather than merely statistically. Mechanical platform improvement is, in Mokyr's terms, prescriptive improvement 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, 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 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 is an instance of Mokyr's pattern: the durable gains came from the extensible, knowledge-intensive layer, not from the bounded mechanical layer. Mokyr's framework also warns that such gains are fragile and that the access cost between research and deployment must stay low for them to persist, which maps onto the question of whether autonomy advances reach the surface through flight-software updates or are frozen at launch.

The two lenses are complementary rather than redundant. Angrist and Pischke tell us how to estimate the relative contribution of the two channels without fooling ourselves; Mokyr tells us how to interpret the estimate once we have it and what it implies for the future trajectory of the technology. 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. Mokyr supplies that account: a wheel is a finished technique whose physics is fully understood and whose marginal returns are therefore decreasing, while autonomy software sits on a research frontier in perception and planning whose propositional base is still widening, so its marginal returns can stay high and its improvements can be retrofitted to deployed systems. Conversely, a historical claim that software is the durable lever would be merely suggestive without a credible estimate that holds terrain and hardware fixed. The dissertation needs both, and the two anchors map cleanly onto the two halves of the work, the estimation design and its interpretation.

---

## 3. Data

### 3.1 Named datasets and access paths

The empirical panel is assembled from three named sources.

1. **PDS rover traverse and localization archives.** The Planetary Data System Geosciences Node hosts the localization and traverse products for the Mars Exploration Rovers, Curiosity (Mars Science Laboratory), and Perseverance (Mars 2020). These archives provide, per drive and per sol, the rover position, the localized traverse path, and the associated mission time, accessible through the PDS Geosciences Node archive interface. Localization within orbital maps, the method that ties surface position to orbital basemaps, is documented in the NTRS record on planetary rover localization within orbital maps (2019). These products are the source of the dependent variable.

2. **NTRS AutoNav and Enhanced Navigation performance reports.** The NASA Technical Reports Server hosts the engineering reports that characterize autonomous-navigation performance by generation, including the Mars Exploration Rover autonomous-navigation results and the Enhanced Navigation description for Perseverance (Verma et al. 2025; Maimone, Cheng, and Matthies 2007). These reports supply the autonomy-generation classification and the autonomous-drive-fraction measures.

3. **TechPort mobility-technology TRL records.** NASA TechPort records the technology-readiness-level history of mobility technologies. 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 here as a robustness check on the autonomy-generation indicators so that the generation classification does not rest only on mission self-description.

### 3.2 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. The panel is therefore an unbalanced panel of drive-sols nested within rovers, spanning the four flight rovers with traverse archives (Sojourner is included where its limited traverse record permits, and is otherwise carried as a boundary case rather than a full panel member because its archive predates the PDS traverse-product standard). A secondary unit, the individual drive, is used for robustness where the archive resolves multiple drives within a sol.

### 3.3 Variable construction

**Dependent variable.** Mobility productivity is operationalized two ways, estimated in parallel. The first is meters traversed per sol, the rate at which the rover converts mission time into distance. The second is its inverse, sols per meter, which is the quantity a planner budgets. Distance is the localized path length from the PDS traverse product, not the straight-line displacement, so that the measure captures the actual path driven rather than net progress.

**Treatment variable.** Autonomy-software generation is a categorical indicator: G1 (Mars Exploration Rover-class AutoNav with visual odometry), G2 (Mars Science Laboratory inherited and extended stack), and G3 (Mars 2020 Enhanced Navigation). A continuous secondary measure is the autonomous-drive fraction, the share of a drive's distance executed under onboard autonomous navigation rather than blind commanded motion, taken from the NTRS performance reports.

**Hardware covariates.** Wheel diameter, mass class, actuator class, and nominal available drive energy per sol, constructed from published mission descriptions (Crisp et al. 2003; Grotzinger et al. 2012; Maki et al. 2020) and TechPort records.

**Terrain covariates.** Terrain class derived from terrain physical properties and slope, drawn from orbital basemaps and from the PDS-archived terrain characterizations (Golombek et al. 2014; Arvidson et al. 2017). Terrain class is the variable that the fixed-effects design absorbs and that the covariate block conditions on.

**Mediator, not control.** Realized slip and realized wheel-soil interaction are recorded but are treated as post-treatment mediators, consistent with the bad-controls discipline (Angrist and Pischke 2009).

### 3.4 Coverage and limitations

Coverage spans the full surface traverse record of the Mars Exploration Rovers, Curiosity, and Perseverance, with Sojourner as a boundary case. The principal limitations are three. 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. Second, autonomy generation and hardware generation are correlated by construction, because each new rover upgraded both at once; this collinearity is the central identification problem and is addressed in Section 4. Third, terrain is not randomly assigned: each rover drove the terrain its landing site presented, so terrain and rover are entangled and must be separated by the terrain-class fixed effects and by within-rover, within-terrain-class comparison.

---

## 4. Research Design and Identification

### 4.1 Estimator and specification

The estimator is a panel linear regression with two-way fixed effects. For drive-sol *i* on rover *r* in terrain class *c*:

```
Productivity_irc = beta1 * AutonomyGen_r
                 + beta2 * Hardware_r
                 + gamma * Terrain_ic
                 + alpha_r      (rover fixed effect)
                 + delta_c      (terrain-class fixed effect)
                 + epsilon_irc
```

Productivity is meters-per-sol in the primary specification and sols-per-meter in a parallel specification. AutonomyGen is the generation indicator (and, in a secondary specification, the continuous autonomous-drive fraction). Hardware is the block of mechanical covariates. Terrain is the continuous terrain block (slope, physical-properties index) that varies within terrain class. The rover fixed effect alpha_r absorbs all fixed, rover-specific characteristics; the terrain-class fixed effect delta_c absorbs systematic differences between terrain classes.

There is a deliberate tension in this specification that the design confronts head-on. Autonomy generation and hardware class are both rover-level attributes, so a rover fixed effect that fully absorbs rover identity would also absorb autonomy generation and leave it unidentified. The design resolves this in two complementary ways. First, the continuous autonomous-drive fraction varies within rover, because the same rover drives some segments fully autonomously and others under blind command depending on terrain and ground decisions; this within-rover variation identifies the autonomy effect even with rover fixed effects in place. Second, for the categorical generation contrast, the design relaxes the rover fixed effect to a terrain-class fixed effect plus the hardware block, so that the generation coefficient is identified from between-rover variation while hardware is explicitly conditioned on. The two approaches bound the effect: the within-rover estimate is conservative and clean, the between-rover estimate is larger in scope but leans on the hardware controls to separate the channels.

### 4.2 Identification strategy

Identification of H1 against H0 rests on the following logic. If productivity gains were purely mechanical (H0), then conditioning on the hardware block should drive the autonomy-generation coefficients to zero and the autonomous-drive-fraction coefficient should vanish once hardware and terrain are controlled. If instead the autonomy channel is real (H1), the autonomous-drive-fraction coefficient should remain positive and significant within rover, where hardware is constant by construction, because the only thing that varies within a rover across autonomous and blind segments is how much the software is doing.

The within-rover autonomous-fraction contrast is the strongest identifying variation available, precisely because it holds hardware fixed at the level of the individual machine. It is the closest thing in this setting to the design-based ideal of comparing like with like (Angrist and Pischke 2014). The between-rover generation contrast is weaker and is reported as a complement, with the hardware block carrying the burden of separating the channels.

A subtle point in the within-rover design deserves to be made explicit, because it is where the bad-controls discipline does its work. The autonomous-drive fraction is chosen by the ground team, not assigned at random: planners drive a segment autonomously when they judge the terrain suitable and the time pressure high, and they drive blind when the terrain is benign or the route is short and well-imaged. This means 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: they condition out the terrain-driven component of the autonomous-fraction choice so that the residual variation in autonomous fraction is closer to as-good-as-random with respect to productivity. What must not be done, and what the design explicitly forbids, is to condition on realized slip or realized drive time, because those are produced by the drive and would absorb the very effect being measured (Angrist and Pischke 2009). The design draws a clean line: a priori terrain and commanded-drive parameters are eligible controls; anything realized during or after the drive is a mediator and stays on the outcome side.

### 4.3 Threats to validity

**Internal validity.** The chief threat is the collinearity of autonomy and hardware generation. The within-rover design mitigates it but cannot fully dissolve the between-rover entanglement. A second threat is bad controls: conditioning on realized slip or realized drive duration would absorb part of the autonomy effect, so these are excluded from the right-hand side and treated as mediators (Angrist and Pischke 2009). A third threat is selection of terrain by mission: planners route each rover along paths chosen partly for their drivability, so observed terrain is endogenous to the rover's capability; terrain-class fixed effects and within-terrain-class comparison address this, but residual selection remains a caveat.

**External validity.** The fleet is four rovers at three landing sites on one planet. Generalization to lunar rovers, to future Mars rovers with different processors, or to off-road terrestrial autonomy is not warranted without re-estimation. The Mokyr lens predicts that 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, but that prediction is itself a hypothesis, not a result.

**Construct validity.** Meters-per-sol is a defensible but partial measure of productivity, because it ignores the scientific value of where the rover went. Autonomy generation is a coarse three-level construct; the continuous autonomous-drive fraction is the finer and preferred measure, with the TechPort TRL records used to check that the generation labels are not merely mission marketing.

**Statistical-conclusion validity.** With few rover clusters, conventional cluster-robust standard errors are unreliable. The design specifies wild-cluster bootstrap inference for the between-rover contrasts and reports the few-cluster caveat explicitly, consistent with the modern fixed-effects literature (de Chaisemartin and D'Haultfoeuille 2020; Goodman-Bacon 2021). Heterogeneity of the autonomy effect across terrain classes is expected and is estimated by interacting autonomy with terrain class rather than assuming a single homogeneous coefficient.

---

## 5. Analysis Plan and Findings

This section is a design-stage analysis plan. The procedures below are specified to be run on the assembled panel. The numbers presented are illustrative expectations stated to make the design concrete, and they are not estimates produced from the data. No empirical result here should be read as a finding. The dissertation is honest on this point: the panel assembly and estimation are the next phase of work, and the contribution at this stage is the design, not the coefficients.

### 5.1 Estimation procedure

1. Assemble the drive-sol panel from PDS traverse products for the Mars Exploration Rovers, Curiosity, and Perseverance, computing localized path length and elapsed sols per drive-sol.
2. Merge the autonomy-generation classification and the autonomous-drive fraction from NTRS performance reports, and merge the hardware covariates and TechPort TRL records.
3. Merge terrain class and terrain covariates from orbital basemaps and PDS terrain characterizations.
4. Estimate the meters-per-sol specification and the sols-per-meter specification in parallel, in three nested forms: terrain fixed effects only; terrain fixed effects plus hardware; terrain fixed effects plus hardware plus autonomy. Compare the incremental explanatory share of the autonomy block against that of the hardware block.
5. Estimate the within-rover autonomous-drive-fraction specification with rover fixed effects, which is the primary test of H1.
6. Compute wild-cluster-bootstrap standard errors for the between-rover contrasts and report the few-cluster caveat.
7. Run the falsification checks of Section 5.3.

### 5.2 Expected or illustrative results (not yet executed)

The following are illustrative expectations under H1, labeled as such.

- **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. Adding the autonomy block is expected to absorb a larger incremental share and to leave the hardware coefficients attenuated, indicating that part of what looked mechanical was the correlated autonomy upgrade.
- **Illustrative, not estimated.** In the within-rover specification, the autonomous-drive-fraction coefficient is expected to be positive and to survive the inclusion of rover and terrain-class fixed effects, because within a single machine the only thing differing between autonomous and blind segments is the software's contribution. This is the result that would most directly support H1.
- **Illustrative, not estimated.** The autonomy effect is expected to be largest in rougher and more 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. This interaction pattern is itself a testable prediction.

If, contrary to these expectations, the autonomy block adds little incremental explanatory share and the within-rover autonomous-fraction coefficient is indistinguishable from zero after hardware and terrain are controlled, H1 is rejected and H0 stands.

### 5.3 Falsification checks

The design specifies in advance the checks that would falsify the contribution. First, if the autonomous-drive-fraction coefficient is zero within rover, the autonomy channel has no support and H1 fails. Second, if the autonomy-generation coefficients lose joint significance the moment hardware enters, the apparent autonomy effect is confounded with hardware and H1 fails. Third, if the terrain-interaction pattern is flat, the mechanism proposed for the autonomy effect (hazard avoidance reduces blind-driving penalty most where hazards are dense) is unsupported even if a raw correlation exists. Naming these checks before estimation is the discipline that makes the contribution falsifiable rather than merely illustrated (Angrist and Pischke 2009).

---

## 6. Discussion

If H1 is supported, the implication for mission design is that mobility productivity is bought more cheaply with software than with mass. A given increment of productivity obtained through better onboard hazard detection and planning costs development and verification effort but little launch mass, and it can in principle be delivered to a rover already on Mars through a flight-software update, as the Mars program has done repeatedly. An increment obtained through a larger wheel or a heavier chassis is paid for at launch and is frozen at the design that left Earth. The Mokyr reading is that the autonomy layer is the extensible, propositional-base-resting layer whose returns are durable, while the mechanical layer delivers real but bounded refinement of a known technique (Mokyr 2002). This would argue for continued investment in onboard autonomy as the primary lever for surface productivity in the Mars Sample Return campaign and in future surface missions, including lunar and crewed-assist rovers.

Rival explanations must be taken seriously. The first rival is that hardware and autonomy are genuinely inseparable in this fleet because they were always upgraded together, so that any attribution to one channel is an artifact of which covariate the analyst chose to privilege. The within-rover autonomous-fraction design is the answer to this rival, because within a single machine the hardware is constant and only the software contribution varies; if that design still shows an effect, the inseparability rival is weakened. The second rival is terrain: later rovers may simply have driven easier ground, so that productivity rose because the terrain got friendlier, not because the rover got smarter. Terrain-class fixed effects and within-terrain-class comparison address this, but residual terrain selection remains a genuine caveat and is acknowledged as such. The third rival is operational learning: ground teams got better at planning drives over twenty years, and that human learning, not onboard autonomy, may drive the trend. Separating onboard autonomy from ground-team learning requires the autonomous-drive-fraction measure, because ground learning improves blind-commanded drives too, whereas the autonomy channel is specifically the share executed onboard.

External validity is bounded. The fleet is small and Mars-specific. The Mokyr framework offers a structured conjecture that the autonomy channel transfers better than the hardware channel, because the underlying perception and planning knowledge is platform-general while a specific actuator is not, but this conjecture would have to be tested on lunar and terrestrial platforms before it could be claimed.

There is a further operational implication worth drawing out, because it connects the statistical question to the way missions are actually run. Surface operations are organized around a daily planning cycle in which the ground team decides how far and where the rover will drive before the next communication window. The binding scarcity in that cycle is sols, and the planning literature treats the embedding of onboard scheduling and decision-making as a way to relax that scarcity (Gao and Chien 2021). 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 that can be trusted to handle hazards autonomously needs less conservative, more sparsely specified commands. That second-order effect is not captured by the meters-per-sol measure alone, and it points to a richer productivity construct for future work, one that counts the ground-team planning effort consumed per meter as well as the sols consumed per meter. The present design deliberately uses 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 ground-team behavior, but the discussion notes that a confirmed autonomy effect would understate the full operational value of the autonomy channel.

A final consideration concerns the asymmetry of the policy 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 errors are not symmetric in their recoverability: a software shortfall can sometimes be patched after landing, while a hardware shortfall cannot. This asymmetry is itself a reason to estimate the two channels carefully rather than to rely on the combined narrative, and it is part of why the contribution is decision-relevant rather than merely descriptive.

What would falsify the contribution is stated plainly and was specified before any estimation: a null within-rover autonomous-fraction coefficient, a collapse of the autonomy-generation block when hardware enters, or a flat terrain-interaction pattern. Any one of these would move the verdict toward H0.

---

## 7. Contribution and Conclusion

This dissertation makes one falsifiable contribution: that cross-rover and within-rover variation in Mars surface mobility productivity is explained better by autonomy-software generation and onboard hazard detection than by mechanical platform improvement, with the null that productivity gains are attributable solely to mechanical improvement. The contribution is delivered as a complete research design rather than as executed estimates, and the document is explicit that the illustrative results are expectations, not findings. The design draws its identification discipline from Angrist and Pischke, in the form of a named counterfactual, a within-unit comparison that holds hardware fixed, a strict bad-controls rule, and few-cluster-aware inference, and it draws its interpretive frame from Mokyr, in the distinction between bounded prescriptive refinement of mechanical hardware and extensible improvement that rests on a deepening propositional base in perception and planning.

The work is decision-relevant because it converts a narrative that credits each rover's success to the whole machine into a design that asks where the next increment of productivity should be bought. If the autonomy channel dominates, the lever for future surface productivity is computational, the cost is paid in development rather than at launch, and the capability can travel to rovers already on the surface. The next phase of work is to assemble the PDS, NTRS, and TechPort panel specified here, to execute the nested and within-rover estimations, to run the three falsification checks, and to report the verdict on H1 against H0 with honest, few-cluster-aware uncertainty.

---

## References

1. Angrist, J. D., and Pischke, J.-S. (2009). *Mostly Harmless Econometrics: An Empiricist's Companion*. Princeton University Press. https://doi.org/10.1515/9781400829828

2. Angrist, J. D., and Pischke, J.-S. (2014). *Mastering 'Metrics: The Path from Cause to Effect*. Princeton University Press.

3. Mokyr, J. (2002). *The Gifts of Athena: Historical Origins of the Knowledge Economy*. Princeton University Press.

4. de Chaisemartin, C., and D'Haultfoeuille, X. (2020). Two-way fixed effects estimators with heterogeneous treatment effects. *Journal of Econometrics*. https://doi.org/10.1016/j.jeconom.2023.105480

5. Goodman-Bacon, A. (2021). Difference-in-differences with variation in treatment timing. *Journal of Econometrics*. https://doi.org/10.1016/j.jeconom.2021.03.014

6. Crisp, J. A., Adler, M., Matijevic, J. R., Squyres, S. W., Arvidson, R. E., and Kass, D. M. (2003). Mars Exploration Rover mission. *Journal of Geophysical Research: Planets*. https://doi.org/10.1029/2002je002038

7. Maimone, M., Cheng, Y., and Matthies, L. (2007). Two years of Visual Odometry on the Mars Exploration Rovers. *Journal of Field Robotics*. https://doi.org/10.1002/rob.20184

8. Helmick, D. M., Cheng, Y., Clouse, D. S., Matthies, L. H., and Roumeliotis, S. I. (2004). Path following using visual odometry for a Mars rover in high-slip environments. *IEEE Aerospace Conference*. https://doi.org/10.1109/aero.2004.1367679

9. Carsten, J., Rankin, A., Ferguson, D., and Stentz, A. (2009). Global planning on the Mars Exploration Rovers: Software integration and surface testing. *Journal of Field Robotics*. https://doi.org/10.1002/rob.20287

10. Carsten, J., Rankin, A., Ferguson, D., and Stentz, A. (2007). Global Path Planning on Board the Mars Exploration Rovers. *IEEE Aerospace Conference*. https://doi.org/10.1109/aero.2007.352683

11. Verma, V., et al. (2025). Enhanced Autonomous Navigation on the Perseverance Mars Rover. *IEEE Transactions on Field Robotics*. https://doi.org/10.1109/tfr.2025.3636366

12. 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*. https://doi.org/10.1007/s11214-020-00765-9

13. 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*. https://doi.org/10.1002/rob.22138

14. 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*. https://doi.org/10.1029/2022je007613

15. Grotzinger, J. P., et al. (2012). Mars Science Laboratory Mission and Science Investigation. *Space Science Reviews*. https://doi.org/10.1007/s11214-012-9892-2

16. 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*. https://doi.org/10.1002/2014je004622

17. Arvidson, R. E., et al. (2017). Mars Science Laboratory Curiosity Rover Megaripple Crossings up to Sol 710 in Gale Crater. *Journal of Field Robotics*. https://doi.org/10.1002/rob.21647

18. 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*. https://doi.org/10.1002/2013je004605

19. Ishigami, G., Miwa, A., Nagatani, K., and Yoshida, K. (2006). Terramechanics-Based Analysis and Traction Control of a Lunar/Planetary Rover. *Field and Service Robotics*. https://doi.org/10.1007/10991459_22

20. Liu, J., et al. (2024). Current-Based Analysis and Validation of a Wheel-Soil Interaction Model for the Trafficability of a Planetary Rover. *Aerospace*. https://doi.org/10.3390/aerospace11110892

21. Gao, Y., and Chien, S. (2021). Autonomy for Space Robots: Past, Present, and Future. *Current Robotics Reports*. https://doi.org/10.1007/s43154-021-00057-2

22. A Comprehensive Review of Path-Planning Algorithms for Planetary Rover Exploration. (2025). *Remote Sensing*. https://doi.org/10.3390/rs17111924

23. Advancements in autonomous mobility of planetary wheeled mobile robots: A review. (2022). *Frontiers in Space Technologies*. https://doi.org/10.3389/frspt.2022.1080291

24. Genova, P., et al. (2013). Simulations of Mars Rover Traverses. *Journal of Field Robotics*. https://doi.org/10.1002/rob.21483

25. Discrete element modelling for wheel-soil interaction and the analysis of the effect of gravity. (2020). *Journal of Terramechanics*. https://doi.org/10.1016/j.jterra.2020.06.002