Robust legged locomotion on granular terrain

Develop robust locomotion methods for legged robots operating on deformable granular terrain, where complex, nonlinear contact forces make balancing and walking substantially harder than on rigid ground.

Background

The paper identifies deformable granular terrain—including sand, gravel, and loose soil—as a particularly challenging environment for field-deployed legged systems. Granular media can behave as either solids or fluids depending on loading conditions, producing difficult-to-model contact forces, sinkage, and foot slip.

The paper addresses this challenge with a three-dimensional resistive force theory contact model and a terrain-adaptive teacher-student reinforcement-learning controller. The reported simulation and hardware results demonstrate progress, but the broader problem of reliable legged locomotion on granular terrain remains unresolved.

References

Locomotion on granular terrain therefore remains a significant open problem for legged robots.

— Learning Terrain-Adaptive Humanoid Locomotion on Granular Terrain  (2609.10286 - Kamohara et al., 9 Sep 2026) in Section I, Introduction

We conjecture that this is induced by the terrain encoder that estiamtes terrain stiffness, and baseline PPO without terrain estimation learns energy consuming gait that works for any terrain condition.

— Learning Terrain-Adaptive Humanoid Locomotion on Granular Terrain  (2609.10286 - Kamohara et al., 9 Sep 2026) in Section V-B, Hardware Experiments