PaperlayerView exact revision on arXiv

Source-linked AI summary

Toward Lunar Legged Robots: Field Deployment Lessons at LUNA

Adrian Fuhrer, Joseph Church, Oliver Fischer, William Talbot, Nicolas Faesch, Yannic Hofmann, Hendrik Kolvenbach, Yusuke Tanaka, Marco Hutter

arXiv:2610.12276v1cs.RO

TL;DR

Lunar legged-robot deployment faces unresolved questions about regolith interaction, dust, and perception under difficult lighting. The paper reports lessons from a 2025 LUNA campaign with ANYmal-D and Magnecko, finding that quadrupeds can traverse regolith simulant while encountering locomotion, perception, and operational limitations. The authors conclude that future development should integrate regolith-aware locomotion, robust perception, repeatable analogue testing, and mission-level validation.

Abstract

from arXiv · show

Legged robots are promising candidates for future lunar surface missions because they can traverse steep, loose, and obstacle-rich terrain that challenges conventional wheeled rovers. However, readiness for lunar deployment is limited by uncertainties in foot-regolith interaction, dust generation, illumination-driven perception degradation, and operational constraints. This paper reports lessons from the 2025 LUNA analogue campaign, where ANYmal-D and Magnecko traversed loose regolith simulant and crater-like terrain and collected long-horizon navigation and visual-inertial data under challenging lighting. We show that quadrupedal robots can traverse regolith simulant, but performance is affected by sinkage and slip, dust-generating contacts, and perception failures caused by overexposure, shadows, and low-texture regions. These results motivate tighter integration of regolith-aware locomotion policies, illumination-robust perception, repeatable analogue testing, and mission-level operational validation for future lunar legged robots.

I. INTRODUCTION

Wheeled rovers are limited on steep, loose, and obstacle-rich terrain, while key questions remain about legged robots’ interaction with lunar regolith and challenging perception conditions. The paper reports 2025 LUNA campaign lessons from two quadrupeds and analyzes system-level challenges relevant to lunar deployment.

I. INTRODUCTION

  • Legged-robot deployment remains uncertain around regolith-aware locomotion, foot traction and sinkage, dust generation, and perception under low texture, strong shadows, and changing illumination.These open questions matter because wheeled rovers have limited traversability in steep, loose, or obstacle-rich terrain.Source: §I · p. 1
  • Lunar analogue campaigns provide more realistic terrain, operational, and integration constraints than typical laboratory settings and can reveal system-level bottlenecks.The paper frames these campaigns as a way to identify locomotion limitations as well as issues involving terrain interaction, perception, navigation, and deployment logistics.Source: §I · p. 1
  • The 2025 campaign operated ANYmal-D and Magnecko in the ESA/DLR LUNA analogue facility under terrain, lighting, sensing, and operational constraints.The observations address how regolith interaction, dust generation, illumination-dependent perception, and field logistics affect readiness for future lunar surface missions.Source: §I · pp. 1–2
  • The paper’s contributions include field deployment observations from two quadrupeds and analysis of challenges in regolith locomotion, perception, dust exposure, and field operations.It also presents lessons and design implications for future lunar legged robots and repeatable analogue testing.Source: §I · p. 2

II. BACKGROUND AND CAMPAIGN CONTEXT

Legged robots may reach steep, irregular, and discontinuous terrain that challenges wheeled rovers, but lunar use raises unresolved questions about granular contact, dust, and perception. The LUNA campaign tested these issues in a controlled indoor facility with prepared regolith terrain and adjustable illumination.

II. BACKGROUND AND CAMPAIGN CONTEXT

  • Legged systems have demonstrated walking, hopping, climbing, and contact-rich mobility for planetary scenarios, including terrain difficult for conventional wheeled rovers,,, [14],, [16],,.For lunar applications, controllers must also address reduced gravity, limited power and thermal budgets, and loose-regolith interaction; simplified contact models leave gaps in understanding granular-terrain behavior and dust generation,.Source: §II · p. 2
  • Lunar analogue campaigns bridge simulation, laboratory testing, and deployment while exposing system-level issues across locomotion, perception, navigation, communication, and operations,,.The LUNA campaign addressed questions about RL-policy regolith interaction, foot-design trade-offs, and perception under low texture, high contrast, and changing illumination.Source: §II · p. 2

A. LUNA Analogue facility and site setup

  • The ESA/DLR LUNA facility provides a controlled indoor environment with EAC-1A simulant and controllable lighting for lunar-mobility testing.The field includes flat regions, craters, and obstacles, and the navigation stack was evaluated under low-angle and high-angle illumination; EAC-1A is dry with very low water content but does not fully represent lunar dust behavior.Source: A. LUNA Analogue facility and site setup · p. 2, Figure 2 · p. 2

B. Hardware setup

Magnecko is a modular quadruped for climbing and rough-surface locomotion, but its joints and torso are not sealed against simulant ingress. A tether-and-spring system combined with controller compensation provided a practical first-order approximation of reduced normal loading during loose-regolith tests.

1) Magnecko:

  • Magnecko is an in-house quadruped developed for climbing and rough-surface locomotion, with modular components and onboard power for untethered operation.Its morphology permits torso-height adjustment for uneven or sloped terrain.Source: 1) Magnecko · p. 2
  • Magnecko’s joints and torso are not sufficiently sealed against simulant ingress, so it was covered with a temporary dust-protection suit.The covering sufficed for short-duration testing, but such coverings can limit heat dissipation and long-term durability; joint durability issues were also discovered on ANYmal-D.Source: 1) Magnecko · p. 2

2) Gravity Offloading:

  • Gravity offloading combined an upward tether force and constant-force spring reel with gravity-compensated feedforward torques to reduce normal loading during loose-regolith locomotion.The approach does not fully reproduce lunar gravity and can introduce tether constraints, but offers a practical first-order test of reduced loading’s effects on sinkage, traction, and locomotion.Source: 2) Gravity Offloading · p. 2

3) ANYmal-D with BOXi:

ANYmal-D carried the BOXi payload for long-horizon navigation testing in the LUNA regolith field, while the campaign also evaluated different foot attachments on both robots.

3) ANYmal-D with BOXi:

  • ANYmal-D carried BOXi, a modular multimodal perception payload for mapping, localization, navigation experiments, and data collection.Fig. 4 shows ANYmal-D equipped with the payload.Source: 3) ANYmal-D with BOXi · p. 2, Figure 4 · p. 3
  • The camera was mounted 70 cm high, walking speed was 0.2–0.6 m/s, and datasets used total-station ground truth.The total station was placed behind the main light source.Source: 3) ANYmal-D with BOXi · p. 3

4) Foot end-effector attachments:

  • Three foot geometries were tested on both robots, including a larger spherical foot and a high-traction design with sawtooth-like teeth.Fig. 6 shows the evaluated foot end-effectors.Source: 4) Foot end-effector attachments · p. 3

C. Reinforcement Learning-Based Locomotion Policy

The campaign used distinct locomotion policies, manually prepared regolith conditions, and challenging lighting to examine locomotion and perception under lunar-analogue constraints.

C. Reinforcement Learning-Based Locomotion Policy

  • Magnecko used a simulation-trained PPO policy with rigid-body Coulomb-friction contacts, which did not explicitly model granular regolith; ANYmal-D used its stock policy.PACE was used for sim-to-real calibration.Source: C. Reinforcement Learning-Based… · p. 3

D. Regolith Simulant Surface Preparation

  • The manually prepared 19 m by 17 m field included loose regolith loosened to about 200 mm and compact/flat regolith transitioning to rocky terrain.These preparations supported evaluation of dust dropout and end-effector contact behavior; Fig. 5 depicts loosened ground and compacted soil.Source: D. Regolith Simulant Surface Preparation · p. 3, Figure 5 · p. 3

E. Illumination setup

  • Raking spotlight-like illumination and, in selected missions, ceiling lights produced strong shadows, low-texture regions, and high-contrast image areas that affected feature tracking and visual localization.The controllable lighting ran at 90–100% intensity and 5600–5900 K.Source: E. Illumination setup · p. 4

IV. CAMPAIGN OBSERVATIONS AND LESSONS LEARNED

Magnecko traversed crater-like loose regolith despite sinkage and slip, while foot geometry showed limited performance differences and ANYmal-D navigation encountered lighting-, dust-, and motion-related perception degradation.

A. Locomotion over Loose Regolith Simulant Crater

  • Magnecko reached the rim of a large crater with its front legs submerged and rear legs slipping, and traversed a mini crater despite traction difficulty.The mini-crater leg struggled across a large height difference under reduced gravity; tether alignment may also have contributed. Figs. 7 and 8 show the crater traversals.Source: A. Locomotion over Loose Regolith… · p. 4, Figure 7 · p. 5, Figure 8 · p. 5
  • Teleoperation commands had to be adjusted to wiggle out of loose regolith, and the policy struggled on the slope because it was not trained with sinkage.The adjustment occurred at t = 13 s in Fig. 7.Source: A. Locomotion over Loose Regolith… · p. 4

B. Foot End-Effector Designs and On-Regolith Simulant Locomotion

  • The three foot geometries showed limited qualitative or quantitative performance differences; regolith packing variations had larger effects, while dust generation differed qualitatively.A larger foot could still submerge in deep loose regolith.Source: B. Foot End-Effector Designs and… · p. 4

C. Long-Horizon Locomotion and Navigation

  • Across 11 missions of 20–100 m traverses, feature tracking was generally stable but weaker in shadows and darker regions, with brief degradation from dust, motion blur, and large interframe displacements.Reported limitations also included overexposure facing the light, low-texture feature loss, dust near sensors, and lower frame rates from onboard compute and real-time processing constraints.Source: C. Long-Horizon Locomotion and… · pp. 4–5

D. Dust, Regolith Disturbance, and Sensor Degradation

The campaign observed dust generated by foot-regolith contacts, with foot motion and end-effector design shaping dust production; the simulant does not fully reproduce lunar dust behavior.

D. Dust, Regolith Disturbance, and Sensor Degradation

  • The controlled dry bed provided system-scale insight, but EAC-1A does not fully reproduce lunar dust behavior under Earth gravity, an atmosphere, and uncharged particles.Source: D. Dust, Regolith Disturbance, and… · p. 5
  • Foot drag and impact were the primary observed mechanisms displacing dust and simulant during legged locomotion.Foot drag can eject particles forward, increasing self-generated dust exposure.Source: D. Dust, Regolith Disturbance, and… · p. 5
  • A more vertical, higher-clearance lift-off may reduce foot-drag dust, but can increase power consumption and cycle time, while touchdown impact also generates dust.Dust-aware locomotion should account for both tangential foot motion and vertical impact velocity.Source: D. Dust, Regolith Disturbance, and… · p. 5
  • High-traction feet with aggressive contact features tended to scoop and eject more regolith than conventional rounded feet.The high-traction design produced visible dust clouds; the effect was less pronounced with rounded designs.Source: D. Dust, Regolith Disturbance, and… · pp. 5–6
  • The campaign recommends treating dust generation as a locomotion constraint and reward when developing future policies and simulations.The proposed aim is to balance traversability, power efficiency, and dust mitigation for long-duration operations.Source: D. Dust, Regolith Disturbance, and… · p. 6

E. Operational Constraints and Field Deployment Lessons

LUNA operations involved facility-specific perception masking, controlled access and communication procedures, and limits on network reliability and terrain repeatability.

E. Operational Constraints and Field Deployment Lessons

  • Indoor walls and structures required environment-specific vision masking so facility features would not enter mapping and navigation pipelines.Masking was needed during operation or post-processing.Source: E. Operational Constraints and Field… · p. 6
  • Controlled hall access, personnel limits, protective equipment, and radio coordination added operational overhead.Clear roles, communication protocols, and training were required before the campaign.Source: E. Operational Constraints and Field… · p. 6
  • Unreliable hall WiFi sometimes required wired LAN, while manual regolith-bed preparation limited terrain repeatability across trials.Local packing variation affected slip, sinkage, and dust generation; careful conditioning and documentation were identified as necessary.Source: E. Operational Constraints and Field… · p. 6

V. REMAINING GAPS TOWARD LUNAR DEPLOYMENT

Remaining deployment gaps include modeling granular contact and dust in locomotion policies, making granular simulation tractable for training, and representing regolith uncertainty in domain randomization.

V. REMAINING GAPS TOWARD LUNAR DEPLOYMENT

  • Lunar legged-robot policies should account for granular contact dynamics, foot sinkage, drag, and dust because these affect robustness, power consumption, exposure, and long-duration operation.Source: A. Regolith- and Dust-Aware Locomotion… · pp. 6–7
  • RL training remains constrained by simplified rigid-body contacts in GPU environments and the computational expense of high-fidelity granular simulation.Open gaps include simulation fidelity, contact-geometry assumptions, and transferability to lunar-relevant regolith.Source: A. Regolith- and Dust-Aware Locomotion… · p. 7
  • Granular domain randomization is difficult because contact parameters depend on foot geometry, impact velocity, gravity direction, and terrain preparation.The paper proposes combining improved granular models, targeted terradynamics experiments, and domain randomization for regolith- and dust-related uncertainty.Source: A. Regolith- and Dust-Aware Locomotion… · p. 7

B. Illumination-Aware Navigation

The paper calls for navigation approaches that account for illumination and sensor degradation, alongside mission-scale verification of route execution, recovery, and scientific data collection.

B. Illumination-Aware Navigation

  • Overexposure, sharp shadows, low-texture terrain, changing appearance, and dust can degrade visual localization, motivating robust exposure control, sensor fusion, and conservative responses to unreliable visual conditions.Future analogue tests should include repeatable lighting and dust-exposure conditions.Source: B. Illumination-Aware Navigation · p. 7

C. Mission-Level Navigation and CONOPS

  • Mission-level navigation and CONOPS need verification beyond short-range locomotion tests, including reliable localization, route execution, operator interaction, and recovery.Future lunar robots may need to operate over sub-kilometer mission scales.Source: C. Mission-Level Navigation and CONOPS · p. 7
  • Scientific measurements may require robots to remain stationary, localize sites, orient instruments, and minimize motion, vibration, or actuator noise during acquisition.Source: C. Mission-Level Navigation and CONOPS · p. 7

D. Toward Higher Technology Readiness

The campaign showed quadrupeds traversing lunar regolith simulant while exposing system-level limitations. Advancing readiness requires lunar-specific design and complementary validation tests.

D. Toward Higher Technology Readiness

  • The tested robots remain at approximately TRL 4–5, and neither ANYmal-D nor Magnecko was designed for planetary surface environments.Higher readiness requires a dedicated lunar design addressing legged locomotion’s power profile, peak actuator loads, and thermal and dust management for exposed leg joints.Source: D. Toward Higher Technology Readiness · pp. 7–8
  • No single terrestrial facility fully replicates the lunar environment, so validation should combine complementary tests from thermal-vacuum locomotion to outdoor mission-level autonomy trials.The proposed tests include LUNA regolith-simulant campaigns, vacuum dust-chamber tests, gravity-offloading experiments, and large-scale outdoor field trials for autonomy and CONOPS.Source: D. Toward Higher Technology Readiness · p. 8

VI. CONCLUSION

  • ANYmal-D and Magnecko traversed loose simulant, crater-like terrain, and obstacle-rich areas, but exposed system-level limitations absent from laboratory tests or rigid-contact simulations alone.These limitations include sinkage, slip, dust generation, illumination-related perception degradation, and field-operation constraints.Source: §VI · p. 8
  • Foot-regolith interaction, illumination, and field operations constrained locomotion, visual localization, and deployment through sinkage, slip, dust, overexposure, shadows, low texture, and operational demands.The paper calls for tighter integration of regolith-aware locomotion policies, illumination-robust perception, repeatable analogue testing, and mission-level operational validation.Source: §VI · p. 8
Loading 2610.12276v1…