cs.ROFeb 21, 2026
SaveScout-Rover cooperation: online terrain strength mapping and traversal risk estimation for planetary-analog explorations
Organizations: University of Southern California, Los Angeles, CA, USA. · University of Pennsylvania, Philadelphia, PA, USA. · NASA Marshall Space Flight Center, Huntsville, AL, USA. · Georgia Institute of Technology, Atlanta, GA, USA. · NASA Ames Research Center
Abstract
Robot-aided exploration of planetary surfaces is essential for understanding geologic processes, yet many scientifically valuable regions, such as Martian dunes and lunar craters, remain hazardous due to loose, deformable regolith. We present a scout-rover cooperation framework that expands safe access to such terrain using a hybrid team of legged and wheeled robots. In our approach, a high-mobility legged robot serves as a mobile scout, using proprioceptive leg-terrain interactions to estimate regolith strength during locomotion and construct spatially resolved terrain maps. These maps are integrated with rover locomotion models to estimate traversal risk and inform path planning. We validate the framework through analogue missions at the NASA Ames Lunar Simulant Testbed and the White Sands Dune Field. Experiments demonstrate (1) online terrain strength mapping from legged locomotion and (2) rover-specific traversal-risk estimation enabling safe navigation to scientific targets. Results show that scout-generated terrain maps reliably capture spatial variability and predict mobility failure modes, allowing risk-aware path planning that avoids hazardous regions. By combining embodied terrain sensing with heterogeneous rover cooperation, this framework enhances operational robustness and expands the reachable science workspace in deformable planetary environments.
Figures & tables
Figure 1: Heterogeneous robot team framework for terrain-aware planetary exploration. (A) A legged scout robot estimates terrain properties using proprioception every step. (B) The terrain properties are used to generate terrain strength and uncertainty maps via kriging. (C) Robot-specific terradynamics models are used to estimate traversal risk for RHex rover and wheeled rover with different morphology and sensing payloads. (D) A central planner assigns mission goals and traversal path based on scientific targets and terrain strength maps, ensuring safe operation.
Figure 2: Experimental setup for terrain-aware legged robot scouting. (A) Experiment site, the SSERVI Regolith Lab testbed located at the NASA Ames Research Center. (B–D) Three terrain-preparation protocols with strength from high to low: tamped (high strength), raked (medium strength), and sifted (low strength). (E)(F) The desired terrain strength map, prepared using the three protocols. (G) Ground truth terrain strength measurements are collected using a robotic penetrometer across the testbed. (H) A legged scout robot performs terrain scouting using proprioception from every step along its path. (I) LiDAR scan after terrain preparation. (J) Penetration resistance measurements from the robotic penetrometer across the testbed. (K) The penetration resistance distribution from terrain grid cells prepared using the three protocols.
Figure 3: Online terrain property mapping during robot traversal. (A) Robot snapshots at successive time steps. (B) Penetration resistance ground truth measured by Traveler. (C) Sequential terrain property maps inferred online by the scout robot, Spirit, from proprioceptive sensing; white markers denote the robot trajectory and inset icons indicate movement direction. (D) Corresponding Gaussian process–based uncertainty maps. (E) Coefficient of determination of the linear terrain mechanics model.
Figure 4: Traversal risk estimation and simulation validation for planetary rovers. (A) Distribution of RHex slip ratio within the mapped region, illustrating increasing traversal risk (red areas) with larger sensor payloads (from 40 to 100 kg). (B) Distribution of slip ratio of the wheeled rover for payload range between 0 kg to 60 kg. (C-E) Simulation validation in a digital twin of the Ames LHS-1 testbed using Chrono. (C, D) RHex (80 kg and 40 kg) moves along a straight path with decreasing strength. The 80 kg robot becomes immobilized in medium-strength terrain (yellow), whereas the 40kg robot successfully traverses the region without failure. (E) RHex (80 kg) moving along a safe path based on the terrain strength map. For all scenarios, the position (middle) and velocity (right) are recorded over time.
Figure 5: Workflow for safe rover mission planning integrating satellite imagery, in-situ measurements, and science-interest-driven goal selection. (A) Overview of the mission plan showing a top-down view of the operational area and the tentative traversal region (purple rectangle). (B) LiDAR-based reconstruction of the traversal region with elevation data (1184–1189 m). (C) LiDAR-based inclination measurements, and initial identification of candidate rover routes. (D) Field interface enabling real-time visualization of terrain strength and science goal selection: (a) estimated toe force over time; (b) toe force during the penetration phase; (c) model fitness; (d) penetration resistance measured from each step; and (e) interpolated penetration resistance map generated via Gaussian Process regression. (E) Field snapshot after Spirit robot scouting. (F) Final penetration resistance map of the scouted region.
Figure 6: Comparison of naive versus safe path planning for rover navigation. (A) Snapshots of rover traversal along the naive path at (1) s, (2) s, (3) s, and (4) s, where the rover speed drops to near zero upon entering a soft region. (B) Snapshots of rover traversal along the safe path at (1) s, (2) s, (3) s, and (4) s, where the rover maintains acceptable forward speed. (C) Predicted rover traversal risk (based on required torque) with the naive path (dashed black) and the safe path (green) overlaid. The color of the path represent experimentally-measured rover speed. (D) Time series of commanded versus actual rover speed during the two traversals. Labels (1)–(4) in A–D correspond to the same waypoints along the path.
Figure 7: Estimating terrain resistance from legged robot interactions using Gaussian Process regression. (A, B) A legged scout collects penetration force and toe position data during each stance phase. From the force-depth profile, terrain resistance is computed. (C) Force and position traces are used to estimate ground contact and penetration depth. (D) Penetration resistance is extracted from the slope of the force-depth curve. Bottom: Gaussian Process regression (kriging) is applied for continuous interpolation of terrain strength. Ground truth (E) shows discrete terrain classes. (F) 1D kriging reconstructs resistance along a path with uncertainty bounds. (G) 2D kriging produces a continuous regolith strength map across the testbed, with uncertainty characterization (H).
Figure 8: Analytical terradynamics models for predicting rover speed. (A) Experimental setup for traversal risk validation using a fluidized granular testbed and a small-scale RHex platform. (B) Terradynamics-based traversal risk model for RHex-class robots [ 47 , 64 ] . (C) Comparison of model-predicted robot speed and experimental measurements. Black markers represent experimentally-measured robot speed for a 0.4 kg rover. Blue, green, orange, red and purple curves represent the model-predicted robot speed for masses of 0.3 kg, 0.4 kg, 0.5 kg, 0.6 kg and 0.7 kg, respectively. (D) Wheeled rover experiments at White Sands analog site. (E) Terradynamics-based traversal risk model for wheeled rovers [ 2 ] . (F) Model-predicted rover slip ratio and axle torque, as a function of terrain penetration resistance per unit area. (G) Traveler to measure the penetration resistance from the fluidized granular testbed. (H) Schematic illustrating the sand solidification depth as a result of comparison between yield force and applied force. (I) Measured terrain penetration resistance per unit area as a function of fluidized bed blower frequency.
| Y/X | 0-1m | 1-2m | 2-3m | 3-4m |
| 0-1m | 4.17 | 1.16 | 0.31 | 0.74 |
| 1-2m | 4.93 | 1.04 | 0.30 | 0.28 |
| 2-3m | 4.46 | 5.75 | 1.41 | 1.65 |
| 3-4m | 1.23 | 4.04 | 3.93 | 1.71 |
| 4-5m | 0.35 | 0.35 | 5.98 | 4.87 |
| 5-6m | 1.03 | 0.56 | 8.38 | 4.16 |
Table 1: Stiffness parameter matrix for NASA Ames LHS-1 simulant testbed. Each cell represents the average stiffness value (N/cm 3 ) for a 1 1 meter region, with coordinates (x, y) corresponding to the physical testbed layout.
| Hyperparameter | Symbol | Value | Bounds | Description |
|---|---|---|---|---|
| Length Scale | 0.5 | |||
| Noise Floor | 0.2 | |||
| Constant | 5 |
Table 2: Gaussian Process Regression (GPR) hyperparameters. Physical descriptions, if applicable, and values of the hyperparameter used a