Terramechanics-Aware Adaptive Quadrupedal Locomotion via Foot-Terrain Interaction Model

Abstract

Locomotion on granular terrain is an essential capability for quadrupedal robots in field applications such as planetary exploration: soft soil yields under load, dissipates propulsive energy, and often alternates abruptly with rigid ground. On such substrates, the mismatch between idealized rigid-contact training and deformable contact can reduce the speed, transport efficiency, and stability of blind reinforcement learning (RL) policies. This paper proposes a lightweight terramechanics-aware adaptive locomotion framework comprising two coupled modules. A virtual ground (VG) module with plastic yield injects realistic foot–terrain forces into standard rigid-contact simulators, capturing nonlinear sinkage, shear saturation, and progressive bearing-capacity loss absent from purely elastic contact formulations. A gait adaptation (GA) module estimates friction and deformability from contact history and aligns gait features with the estimated terrain properties, promoting terrain-specific gaits without exteroception. Experiments across Isaac Sim, MPM-based Newton Physics simulation, and Unitree Go2 hardware show that VG improves granular-terrain mobility and contact regularity, while GA enables stable online gait switching across rigid–soft transitions. Indoor loose-sand experiments demonstrate shallower, more regular footprints with VG and adaptive gait changes during a sand-to-tile transition; field trials on a natural sandy beach demonstrate more efficient sustained mobility, with higher sustained speed and lower energy cost than blind-RL baselines.