--- marp: true theme: default paginate: true --- # Surface Mobility Productivity Across the Mars Rover Fleet ## What drives drive-distance and sols-per-meter? Dissertation Defense Brief Candidate JPL_AUTONOMY_EDL_03 COLLEGIUM 1st Battalion NORTH STAR / JPL category: Autonomous Systems and Robotics Methodological anchors: Angrist and Pischke; Mokyr 2026-06-15 Design-stage dissertation. No executed empirical results. --- # The contribution One falsifiable proposition, stated before any estimation. - **H1 (contribution):** Conditional on terrain class and the same hardware covariates, autonomy-software generation and onboard hazard detection explain the **larger** share of between-rover and within-rover variation in per-sol mobility productivity; the autonomy coefficients are jointly significant and remain so after hardware enters. - **H0 (null):** Conditional on terrain class, that variation is explained by mechanical platform covariates and **not** by autonomy-software generation; the autonomy coefficients are jointly indistinguishable from zero once hardware is included. The deliverable is a complete, pre-registered design that can reject H1, not a set of estimates. --- # Why it matters, stated up front - Mobility is the binding constraint on a rover's science return. Sols spent driving are sols not spent sampling, placing instruments, or caching. - Mobility productivity rose steeply across the fleet: from about one hundred meters for Sojourner to multi-kilometer traverses and single-sol distance records for Perseverance. - The two candidate causes, software and steel, are bought from different budgets, paid at different points in the lifecycle, and recoverable to different degrees. - The policy error is **asymmetric**: a software shortfall can sometimes be patched after landing; a hardware shortfall cannot. --- # The problem frame - **Current state:** Each rover is described as both a better machine and a smarter one; the gain is credited to the mission as a whole (Crisp et al. 2003; Grotzinger et al. 2012; Verma et al. 2025). - **Desired state:** A design that holds terrain and hardware fixed and quantifies how much of the gain is the autonomy channel alone. - **Gap:** No identification strategy separates the channels; they were upgraded together every mission and are collinear by construction. - **Consequence:** Planners cannot rationally allocate a fixed mass-and-power budget between heavier hardware and more onboard computation. --- # The literature gap - The space-robotics literature is rich in description and thin in identification. - It supplies the dependent variable (verified localized path lengths), the treatment construct (G1/G2/G3), the covariates (terramechanics), and the within-machine mode variation the design needs. - It never builds a comparison that holds terrain and hardware fixed and partitions the productivity gain. - 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 and never identified (Gao and Chien 2021). --- # Theoretical framework: two anchors, non-redundant - **Angrist and Pischke (design-based econometrics):** governs estimation. Name the counterfactual; compare like with like; forbid bad controls; be honest about few clusters; specify in advance what would refute the claim. - **Mokyr (economic history of technology):** governs interpretation. A wheel is a finished technique with bounded, decreasing returns; autonomy software rests on a widening propositional base and is extensible and retrofittable. - The estimation lens recovers a contribution it cannot interpret. The interpretation lens predicts a pattern it cannot establish. The design is the join. --- # Data: three public, mission-external archives - **PDS rover traverse and localization archives** (MER, MSL, Mars 2020): per-sol localized path length and elapsed sols. Source of the dependent variable. - **NTRS AutoNav and Enhanced Navigation reports:** autonomy-generation classification and the autonomous-drive fraction. - **TechPort mobility-technology TRL records:** external check that G1, G2, G3 are genuinely distinct generations, not mission self-description. - Unit of analysis: the **drive-sol**, nested within rover. Unbalanced panel; Sojourner is a qualitative boundary case (predates the traverse-product standard). --- # Key variable definitions - **Dependent variable:** meters per sol and its inverse, sols per meter. Distance is the **localized path length**, not straight-line displacement. - **Treatment:** autonomy-software generation (categorical G1/G2/G3) and the continuous **autonomous-drive fraction** (share of a drive executed under onboard autonomy rather than blind command). - **Hardware covariates:** wheel diameter, mass class, actuator class, available drive energy per sol. - **Terrain:** terrain class (the fixed-effect dimension) plus slope and a physical-properties index. - **Mediators, NOT controls:** realized slip and realized wheel-soil interaction stay on the outcome side. --- # Design and identification Two-way fixed-effects panel regression: ``` Productivity_irc = beta1*AutonomyGen_r + beta2*Hardware_r + gamma*Terrain_ic + alpha_r (rover FE) + delta_c (terrain-class FE) + epsilon_irc ``` - Estimated both ways: meters-per-sol and sols-per-meter. - **Primary test:** the within-rover autonomous-drive-fraction coefficient, which holds the mechanical platform fixed at the level of one machine. - **Complement:** the between-rover generation contrast, with the hardware block conditioned on. --- # Why the within-rover contrast is the load-bearing identification - Autonomy generation and hardware class are both rover-level and nearly collinear; a rover fixed effect that absorbs identity also absorbs the generation. - Within a single machine the wheels, mass, actuators, and energy are constant by construction. - The only thing that varies between an autonomous-heavy and a blind-commanded segment is how much the software is doing. - This is the closest the fleet permits to the design-based ideal of comparing like with like (Angrist and Pischke 2014; Biesiadecki et al. 2007). --- # The bad-controls rule - Realized slip, realized drive time, and realized wheel-soil interaction are produced by the drive, partly by how the autonomy software chose to drive. - Conditioning on them would absorb part of the effect and bias the estimate toward the null, manufacturing a false rejection of H1. - A priori terrain and commanded-drive parameters are eligible controls; anything realized during or after the drive is a mediator. - The line is underwritten twice: by the econometrics (Angrist and Pischke 2009) and by terramechanics, which models slip as downstream of machine, soil, and commanded motion (Ishigami et al. 2007; Ding et al. 2009). --- # The named causal mechanism Not a correlation. Each link is a specific transmission the design can examine. ``` Driver: advance in autonomy generation (G1 -> G2 -> G3), culminating in Enhanced Navigation that plans while driving Mechanism: a larger share of each drive runs under onboard hazard detection and real-time planning rather than blind commanded motion Effect: higher autonomous-drive fraction and more meters per sol, concentrated in hazard-dense terrain Operational: fewer sols per meter, freeing sols for sampling and caching Strategic: productivity bought with software, paid in development not at launch, uploadable to a rover already on the surface ``` --- # Threats to validity - **Internal:** autonomy and hardware upgraded together. Within-rover design holds hardware fixed; bad controls excluded; terrain-class FE for selection; autonomous-fraction measure separates onboard autonomy from ground-team learning. - **External:** four rovers, three sites, one planet. No automatic transfer to lunar or terrestrial platforms. - **Construct:** meters-per-sol ignores science value; the autonomous-drive fraction is the finer measure; TechPort checks the generation labels. - **Statistical-conclusion:** few clusters; wild-cluster bootstrap; heterogeneity by terrain estimated, not assumed. --- # Direction of the residual bias - The autonomous fraction is chosen by the ground team, not assigned at random, and is correlated with terrain. - Teams authorize more autonomy on harder, hazard-dense ground and drive blind on benign flats. - Imperfect terrain conditioning therefore makes the autonomous fraction proxy partly for **difficult, low-productivity** terrain. - That biases the within-rover coefficient **downward**, toward zero. A positive, significant result is conservative with respect to this threat. --- # Analysis plan, pre-registered 1. Assemble the drive-sol panel from PDS traverse products. 2. Merge autonomy generation, autonomous-drive fraction, hardware, terrain. 3. Nested decomposition: terrain; terrain + hardware; terrain + hardware + autonomy. Compare incremental explanatory share. 4. Estimate the within-rover autonomous-fraction specification (primary test of H1). 5. Wild-cluster bootstrap inference for the few-cluster between-rover contrasts. 6. Run the three falsification checks, placed last so they cannot tune the specifications. --- # The fixed decision rule A conservative conjunction. All three must hold to sustain H1. 1. The within-rover autonomous-fraction coefficient is positive (meters-per-sol), distinguishable from zero under wild-cluster bootstrap, and survives leave-one-rover-out and measurement-error bounding. 2. In the nested decomposition the autonomy block adds a larger incremental share than hardware, and hardware coefficients attenuate when autonomy enters. 3. The terrain-interaction pattern is not flat: the effect is larger in hazard-dense terrain. Any one failing moves the verdict toward H0. --- # Expected results (illustrative, design-stage, not executed) Labeled expectations deduced from the mechanism, never estimates from data. - Hardware block alone absorbs a moderate share of between-rover variance (real mechanical gains). - Adding the autonomy block absorbs a **larger** incremental share and attenuates the hardware coefficients. - The within-rover autonomous-fraction coefficient stays positive after rover and terrain fixed effects. - The autonomy effect is largest in rough, hazard-dense terrain and smallest on benign flats. The Chapter 6 result tables are specified but deliberately **unpopulated by design**. --- # Threats to the verdict and what would falsify the contribution Specified before any estimation: 1. A null within-rover autonomous-fraction coefficient after terrain conditioning. 2. The autonomy-generation block losing joint significance the moment hardware enters. 3. A flat terrain-interaction pattern, which would undercut the proposed mechanism even if a raw correlation survived. Naming these in advance is what makes the contribution falsifiable rather than illustrated. --- # Rival explanations, engaged - **Inseparability** (channels always upgraded together): largely defeated for the marginal within-machine claim, because within a rover only the software varies. - **Terrain selection** (later rovers drove easier ground): mitigated by terrain-class fixed effects; residual within-class selection survives as the **standing caveat**. - **Ground-team operational learning** (people improved, not the machine): differenced out under a symmetry assumption, because learning improves blind and autonomous segments alike; asymmetric learning is a flagged refinement. The design names which rival it cannot fully dispatch rather than claiming to have killed all three. --- # Power and the limits of a small panel - **Within-rover contrast:** comparatively well-powered; identifying variation is hundreds to thousands of drive-sols per machine. Binding limit is measurement error in the reconstructed autonomous fraction, not cluster count. - **Between-rover contrast:** structurally underpowered with three clusters; reported as suggestive, never as a confirmatory basis for H1. - A null on the well-powered within-rover test is meaningful evidence against H1; a null on the under-powered between-rover test is weak evidence about anything. The design weights them accordingly. --- # Confidence and uncertainty, calibrated - **That productivity rose and the channels were never separated:** high confidence, on the convergent mission and survey record. - **That sols are the binding scarcity that makes the decomposition decision-relevant:** high confidence, on the mission-campaign record. - **That the within-rover contrast recovers the mechanism:** moderate to high; conditional mean independence is an assumption the design makes defensible but cannot prove. - **That the design improves on the alternatives:** high; no valid instrument exists, and the within-rover panel is the cleanest comparison the fleet permits. - **That the remaining identification risk is bounded:** moderate; three rovers will never be many. - Highest-value gap to close before execution: a clean per-drive autonomous-fraction series across all generations. --- # How the argument holds together **The whole design exists to make one question adjudicable:** a falsifiable design can decide H1 versus H0, holding terrain and hardware fixed, and be rejected if the autonomy block fails once hardware enters. | Claim | Where it is made | Confidence | |---|---|---| | Productivity rose and the channels were never separated | Ch 1, 3 (mission + survey record) | High | | Sols are the binding scarcity that makes the decomposition matter | Ch 1, 7 (sols, MSR, asymmetric error) | High | | The within-rover contrast recovers the mechanism | Ch 4, 5 (within-rover + bad-controls) | Mod-high | | The within-rover design improves on the alternatives | Ch 5, 6 (vs naive between-rover) | Mod-high | | The remaining identification risk is bounded | Ch 5, 6 (few-cluster + falsification) | Moderate | This is a research design over public archives, not an architecture or system build. --- # Residual risk, stated plainly - Autonomous-drive fraction is partly reconstructed; classical measurement error attenuates the within-rover coefficient toward the null. Read conservatively. - Residual within-terrain-class selection survives after coarse, constructed terrain classes. - The between-rover generation contrast is underpowered with three clusters. - External validity is bounded to the Mars fleet; lunar and terrestrial transfer is a Mokyr-structured conjecture, not a result. Each is named, signed in direction where possible, and priced into the confidence above. --- # What stands regardless of the verdict - The **separation** of two causes the literature always reported fused. - The **asymmetric-error argument** that makes the separation worth doing, logically prior to any coefficient. - The reusable **measurement apparatus** (drive-sol panel tied to generation, fraction, hardware, terrain). - The **interpretive frame** that converts any estimate into a statement about whether the future lever is the extensible software layer or the bounded, launch-frozen mechanical one. - A credible null from a powered design is itself a finding (Angrist and Pischke 2009). --- # Contribution restated - One falsifiable claim: autonomy-software generation, not the mechanical platform, is the dominant driver of cross-rover mobility productivity, with a pre-specified rule that can reject it. - Decision-relevant: if software dominates, the lever is computational, paid in development not at launch, and uploadable to a rover already on Mars. - Directly relevant to Mars Sample Return caching and to future surface and crewed-assist rovers. - The next phase executes the design on the full PDS / NTRS / TechPort panel and reports the verdict with few-cluster-aware uncertainty. --- # Defense questions anticipated - Can you separate autonomy from hardware when every rover upgraded both? Show the within-rover variation. - Is the autonomous-drive fraction endogenous to terrain? How do the terrain fixed effects address it, and which way does residual selection bias the estimate? - Why is realized slip a bad control rather than a useful one? - With three clusters, what does inference even mean here? - Does ground-team learning, not onboard autonomy, drive the trend? - What single result would make you abandon H1? --- # References (selected anchors and primary sources) - Angrist, J. D., and Pischke, J.-S. (2009). *Mostly Harmless Econometrics*. doi:10.1515/9781400829828 - Angrist, J. D., and Pischke, J.-S. (2010). The Credibility Revolution in Empirical Economics. *JEP*. doi:10.1257/jep.24.2.3 - Mokyr, J. (2002). *The Gifts of Athena*. Princeton University Press. - Crisp et al. (2003). Mars Exploration Rover mission. doi:10.1029/2002je002038 - Maimone, Cheng, Matthies (2007). Two years of Visual Odometry on the MERs. doi:10.1002/rob.20184 - Biesiadecki, Leger, Maimone (2007). Tradeoffs Between Directed and Autonomous Driving. doi:10.1177/0278364907073777 - Verma et al. (2025). Enhanced Autonomous Navigation on Perseverance. doi:10.1109/tfr.2025.3636366 - Gao and Chien (2021). Autonomy for Space Robots. doi:10.1007/s43154-021-00057-2 Full 128-entry reference list appears in the dissertation back matter.