Locomotion analysis of a quadruped interacting with the lunar granular surface

Technical Summary: Locomotion Analysis of a Quadruped Interacting with the Lunar Granular Surface

Problem Statement
Legged robots are increasingly proposed for extra-terrestrial exploration, particularly for navigating complex terrains like lunar pits and lava tubes where wheeled rovers struggle. However, deploying these robots on the Moon presents unique challenges, including energy constraints, thermal management, and the specific nature of the lunar surface. The lunar regolith is a granular, sand-like substrate that behaves differently than the rigid ground assumptions often used in locomotion control. Current Reinforcement Learning (RL) policies trained with rigid contact models (e.g., Coulomb friction) may fail to generalize to soft, deformable surfaces where soil flow and viscous drag occur beneath the feet. This mismatch risks instability, poor tracking, and excessive energy expenditure, which are critical failures given the limited power budget of lunar missions.

Methodology
The study implements a terradynamic model of robot foot-granular surface interactions within a Reinforcement Learning simulation environment to analyze locomotion performance.

• Terradynamic Modeling: The authors integrate the contact model proposed by Li et al. [15], which treats the lunar surface as a viscous granular medium. Instead of simple friction, the model calculates lift and drag forces ($F_z, F_x$) based on the integration of stresses over the foot's contact area. These stresses are functions of penetration depth, angle of attack, and angle of intrusion, parameterized by Fourier coefficients and a scaling factor ($\zeta$) representing soil compactness and grain size. The model is discretized over a mesh representation of the foot, allowing for generalization across different foot morphologies.
• Simulation Environment: The simulation was conducted in RaiSim. Two environments were compared:
  1. Rigid Contact: Standard Coulomb friction with a spherical foot model.
  2. Soft Contact: The terradynamic model with a hemispheric mesh foot (688 faces) intruding into a virtual soil layer ($z_H = 0.05$m).
• Robot and Control: A 13.5 kg quadruped (analogous to the Magnecko robot with an Arachnid hip configuration) was used. Locomotion was controlled via a Deep Reinforcement Learning policy trained using Proximal Policy Optimization (PPO). The policy maps observations (base velocity, joint states, gravity projection, and velocity commands) to target joint angles.
• Training Strategy: To address the instability of training from scratch in soft environments, transfer learning was employed. A policy trained for 800 episodes on rigid terrain served as the initialization for training in the soft terrain environment. Curriculum learning was used to gradually increase terrain difficulty (slope and roughness).
• Performance Metrics: The study analyzed gait characteristics qualitatively and quantitatively measured joint torques, mechanical power, power losses (accounting for motor, gear, and winding inefficiencies), total energy expenditure, and the Mechanical Cost of Transport (MCOT).

Key Results
Comparing the policies trained on rigid versus soft contact environments revealed significant differences in locomotion behavior and efficiency:

• Gait Characteristics: The soft contact environment necessitated a qualitatively different gait. The robot adopted smaller strides with a higher duty cycle to minimize slip forces, whereas the rigid contact gait utilized wider strides relying on high-impulse reactions. The soft contact gait exhibited more sideslip and reduced tracking accuracy compared to the rigid baseline.
• Torque and Power:
  • Joint Torques: Maximum joint torques were generally lower in the soft terrain due to the lower impulse of the contact forces compared to rigid impacts.
  • Mechanical Power: Counter-intuitively, the mechanical power output required to move the robot was lower in the soft terrain.
  • Power Loss and Total Energy: Despite lower mechanical power, the total energy expenditure was higher in the soft terrain. This is attributed to sustained contact forces during both the stance and swing phases, leading to increased power losses in the motor and transmission systems. The "soft" interaction requires continuous effort to push through the granular medium, resulting in greater energy dissipation over the same distance.
• Stability and Outliers: The soft contact simulations showed higher variance in total energy and power, with maximum values significantly exceeding rigid contact cases. This indicates instances where the robot became stuck in the regolith or experienced unstable dynamics, requiring high power to recover.

Significance and Claims
The paper claims that the analysis demonstrates the necessity of incorporating granular terrain modeling into the co-design of mechanical systems and locomotion controllers for extra-terrestrial legged robots.

• Impact on Design: The study asserts that assuming rigid contacts can lead to suboptimal or unsafe designs. The specific interaction dynamics of granular surfaces (viscosity, intrusion/extrusion forces) fundamentally alter the optimal gait and energy requirements.
• Control Implications: RL policies trained solely on rigid assumptions are insufficient for lunar regolith. The research highlights that while a common set of reward weights can stabilize training for both environments, the resulting gaits are distinct.
• Future Relevance: The authors conclude that understanding foot-terrain interactions for granular substrates is imperative for selecting appropriate leg morphologies, motor drives, and control strategies to ensure robust and energy-efficient traversal of lunar surfaces, particularly for missions targeting difficult-to-reach features like lava tubes.

The work does not claim to have solved all sim-to-real gaps but establishes that ignoring the viscous nature of granular media leads to inaccurate predictions of energy consumption and gait stability, which are critical constraints for lunar exploration missions.