cs.ROOct 28, 2026

Explore Where It Matters: Targeted Exploration for Efficient Rover Adaptation

Authors: Jacob Levy, Arthur Bouton, Tristan D. Hassler, Hari Nayar, David Fridovich-Keil

Organizations: Department of Aerospace Engineering, University of Texas at Austin · Jet Propulsion Laboratory, California Institute of Technology

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Future planetary rovers will traverse steep, deformable terrain where slip, sinkage, and terrain deformation can threaten mobility. Although simulation can capture some wheel-soil effects, unmodeled deformable-terrain phenomena create sim-to-real gaps requiring deployment-time adaptation. Yet, rover operations leave little room for extensive trial-and-error. We introduce TERRA (Targeted Exploration for Reward-Relevant Adaptation), an online adaptation framework that efficiently seeks deployment data exposing task-relevant model errors and corrects them. TERRA adapts a simulation-trained world model with a reward-weighted prediction loss and trains a loss predictor to estimate where future performance-relevant errors are likely to occur, using this signal to guide exploration during MPC. In simulation, TERRA reaches approximately 90% slope-climbing success within five trials under severe out-of-distribution pretraining conditions, outperforming exploration baselines in reward, completion time, and energy to first success. On the ERNEST rover, TERRA recovers coordinated active-suspension locomotion on a deformable Mars-simulant slope within three to five real-world trials. Project website: https://terra-explore.github.io

Figures & tables

Figure 1: TERRA rapidly adapts to an out-of-distribution deformable slope. No-Slope pretraining never includes slopes, while Hard-Ground never requires coordinated use of the Active Gimbal Suspension (AGS). Both initially entrap the ERNEST rover, but TERRA learns to climb using its AGS within three to five trials.
Figure 2: AGS actuation in sim.
Figure 3: TERRA overview. A simulation-pretrained world model rolls out candidate action sequences for sampling-based planning, while a loss predictor estimates where future performance-relevant dynamics errors are likely to occur. The planner combines predicted task reward with this targeted exploration signal, executes the selected action on the rover, and updates both models with the resulting deployment data.
Figure 4: Simulated slope-climbing adaptation under two severe out-of-distribution pretraining conditions. Across 100 random seeds, TERRA reaches approximately 90% success within five deployment trials in both No-Slope and Hard-Ground, while achieving higher task reward, shorter completion times, and lower energy to first success than exploration baselines. Curves show means with 95% bootstrapped confidence intervals.
 No-SlopeHard-Ground
 SR ↑\uparrowEnergy (kJ) ↓\downarrowSR ↑\uparrowEnergy (kJ) ↓\downarrow
TERRA0.9150.3 [47, 53]0.8951.6 [49, 54]
No expl.0.8556.5 [52, 61]0.8561.0 [57, 65]
No RWL0.7380.9 [74, 88]0.8660.4 [56, 65]
No expl. / RWL0.6178.7 [72, 86]0.8757.3 [53, 62]
Table 1: Ablations without exploration (expl.) and/or reward-weighted loss (RWL). Across 100 seeds, we report success rate (SR) at trial 5 and energy to first success with 95% bootstrapped CIs.
Figure 5: Adaptation metrics for Hard-Ground. Left: each method’s exploration signal evolution over deployment. Middle: lower prediction error alone does not necessarily yield better task performance. Trial-5 error and reward-gradient magnitude for TERRA and RND, showing that TERRA reduces error where the reward is most sensitive, while RND achieves lower aggregate error largely on reward-insensitive observation groups.
Figure 6: Real-world AGS yaw during Hard-Ground. TERRA adapts from little AGS use in Trial 1 to large alternating yaw motions by the successful Trial 3 climb.
Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

ObservationDimension
Pose, SE(2)3
AGS Yaw and Roll2
Body Roll and Pitch2
Wheel Forces12
Wheel Torques12
Height Map Sampling210
Appendix
Table 2: Observation space.
ActionDimension
Commanded Forward Velocity1
AGS Yaw Position Command1
AGS Roll Position Command1
Appendix
Table 3: Action space.
Term NameReward TermWeight
Forward advance reward−xt+1−xtΔt\displaystyle\phantom{-}\dfrac{x_{t+1}-x_{t}}{\Delta t}50.050.0
Lateral deviation penalty−yt+12\displaystyle-y_{t+1}^{2}2.52.5
Bearing deviation penalty−βt+12\displaystyle-\beta_{t+1}^{2}1.01.0
AGS Yaw penalty−qyaw,t+12\displaystyle-q_{\mathrm{yaw},t+1}^{2}1.251.25
AGS Roll penalty−qroll,t+12\displaystyle-q_{\mathrm{roll},t+1}^{2}10.010.0
Yaw action penalty−ayaw,t2\displaystyle-a_{\mathrm{yaw},t}^{2}1.01.0
Appendix
Table 4: Task reward function, RtaskR^{\mathtt{task}} . Here, xx and yy are the forward and lateral displacements of the robot; β\beta is the deviation from initial robot bearing; and qyawq_{\mathrm{yaw}} and qrollq_{\mathrm{roll}} are the AGS joint angles.
ParameterValue
Discount factor0.99
Soft target update factor5\times10−35\text{\times}{10}^{-3}
Policy update delay2
Target policy regularization noiseN(0, 0.052)\mathcal{N}(0,\,0.05^{2})
Bounds of the target policy regularization0.2
Appendix
Table 5: Hyperparameters for RL with TD3.
Pretraining ConditionRipplesStepFlatSlope
No-Slope10241024––
Hard-Ground512–512512
Appendix
Table 6: Number of parallel environments for each pretraining condition.
ParameterValue / Range
Initial rover pose
Forward position ( m\mathrm{m} )[−1.0, 0.0][-1.0,\,0.0]
Lateral position ( m\mathrm{m} )[−0.2, 0.2][-0.2,\,0.2]
Bearing angle ( \SIUnitSymbolDegree\mathrm{\SIUnitSymbolDegree} )[−5.0, 5.0][-5.0,\,5.0]
Soil Parameters
Soil density ( kg m−3\mathrm{kg}\text{\,}{\mathrm{m}}^{-3} )16001600
Appendix
Table 7: Common parameters during RL for both the No-Slope and Hard-Ground pretraining conditions. Individual values indicate fixed parameters and ranges indicate uniform random sampling within the provided ranges.
ParameterValue / Range
Soil Parameters
Internal friction angle ( \SIUnitSymbolDegree\mathrm{\SIUnitSymbolDegree} )[30, 40][30,\,40]
Step
Episode length ( s\mathrm{s} )4545
Height ( m\mathrm{m} )[0.2, 0.4][0.2,\,0.4]
Roll ( \SIUnitSymbolDegree\mathrm{\SIUnitSymbolDegree} )[−1, 1][-1,\,1]
Appendix
Table 8: Parameters during RL for No-Slope only. Individual values indicate fixed parameters and ranges indicate uniform random sampling within the provided ranges.
ParameterValue / Range
Soil Parameters
Internal friction angle ( \SIUnitSymbolDegree\mathrm{\SIUnitSymbolDegree} )6060
Flat
Episode length ( s\mathrm{s} )3030
Roll ( \SIUnitSymbolDegree\mathrm{\SIUnitSymbolDegree} )[−5, 5][-5,\,5]
Pitch ( \SIUnitSymbolDegree\mathrm{\SIUnitSymbolDegree} )[−5, 5][-5,\,5]
Appendix
Table 9: Parameters during RL for Hard-Ground only. Individual values indicate fixed parameters and ranges indicate uniform random sampling within the provided ranges.
ParameterValue
Embedding dimension128128
Encoder (observations) MLP hidden dims256,256256,256
All transformers MLP hidden size512512
Dynamics transformer layers44
Dynamics transformer heads88
Base policy transformer layers44
Appendix
Table 10: World model and loss predictor architectural parameters.
Figure 7: Detailed world model and loss predictor architecture.
ParameterValue
Candidate actions batch size10241024
Noised base policy actions batch size22
Solver iterations1010
Initial action standard deviation[0.2,0.7,0.78][0.2,0.7,0.78]
Minimum action standard deviation0.0010.001
Base policy action standard deviation[0.025,0.05,0.05][0.025,0.05,0.05]
Appendix
Table 11: MPPI parameters.
ParameterValue
LoRA rank6464
Ordinary prediction error weight, c1c_{1}0.040.04
Reward-weighted prediction error weight, c2c_{2}0.0050.005
Loss bin thresholds, {τj}j=1N\{\tau_{j}\}_{j=1}^{N}{0,0.05,0.1,0.2,0.4,0.8}\{0,0.05,0.1,0.2,0.4,0.8\}
Class weights, ww[0.5,0.75,1,5,10,20][0.5,0.75,1,5,10,20]
Exploration reward strength, λ\lambda0.80.8
Appendix
Table 12: TERRA parameters.
 SR ↑\uparrowReward ↑\uparrowTime ( s\mathrm{s} ) ↓\downarrowEnergy ( kJ\mathrm{kJ} ) ↓\downarrowREL ↑\uparrow
TERRA0.912.46 [2.3, 2.6]83.4 [82, 84]50.3 [47, 53]1.00
(E.1) Adding Reward-Weighted Adaptation to Baselines
RND0.882.46 [2.3, 2.6]84.5 [83, 86]53.1 [50, 57]0.98
P2E0.882.36 [2.2, 2.5]84.8 [83, 86]52.4 [50, 55]0.97
RUNE0.811.98 [1.8, 2.2]87.3 [86, 89]52.8 [50, 56]0.90
Disagreement0.802.32 [2.1, 2.5]90.5 [88, 93]59.2 [56, 63]0.90
Appendix
Table 13: Additional experiments and ablations. Unless noted otherwise, No-Slope, Trial-5, 100 seeds; brackets show 95% bootstrap CIs. REL is the geometric mean performance ratio to TERRA across success rate, reward, completion time, and energy to first success; REL<1\text{REL}<1 is worse than TERRA and REL>1\text{REL}>1 is better.
More in this topic

Planetary Rover Robotics

See all 20 papers

Explore similar work

CardsList

  1. Learning All-Terrain Locomotion for a Planetary Rover with Actively Articulated Suspension

    Jun 5, 2026Arthur Bouton, Tristan D. Hasseler, Michael Paton +5Planetary Rover RoboticsOff-Road Robot Navigation

    Jet Propulsion Laboratory, California Institute of Technology, Pasadena, CA 91109 USA · Center for Autonomy, University of Texas at Austin, Austin, TX 78712 USA · Space Systems Laboratory, University of Maryland, MD 20742 USA

  2. Proprioceptive Safe Active Navigation and Exploration for Planetary Environments

    Mar 9, 2026Matthew Y. Jiang, Feifei Qian, Shipeng LiuPlanetary Rover RoboticsAutonomous Robot Exploration

    Georgia Institute of Technology, Atlanta, GA 30332, USA · University of Southern California, Los Angeles, CA 90089, USA

  3. Terrain-Adaptive Grouser Wheel for Optimal Planetary Exploration: Design and Experimental Investigation

    May 23, 2026Vincent Griffo, Yashwanth Kumar NakkaGranular Terrain LocomotionRobotics

    Aerospace Robotics Lab, Daniel Guggenheim School of Aerospace Engineering, Georgia Institute of Technology, Atlanta, GA, USA