Teaching a Humanoid to Walk on Shifting Ground

Our new paper, Learning Terrain-Adaptive Humanoid Locomotion on Granular Terrain, is now on arXiv.

Basalt locomotion clip from the project page.

The ground is part of the dynamics

A foot can sink into sand and experience drag as it moves through the material. Training only with rigid contact misses these interactions. We use three-dimensional resistive force theory (3D RFT) to model granular contact efficiently enough for reinforcement learning in IsaacLab.

Learning to recognize the terrain

A teacher learns with privileged terrain information and a variational autoencoder that represents material properties. A student then learns from the teacher using proprioceptive history, allowing deployment without those privileged measurements. The representation helps the controller adjust its gait to the terrain beneath it.

From simulation to hardware

We evaluate transfer in Newton’s material point method (MPM) simulator and on real basalt, dry sand, and beach sand. These tests examine both the contact model and the controller’s ability to adapt beyond its training environment.

For me, the central lesson is that contact modeling and adaptation belong together: the simulator needs to expose useful physical variation, and the policy needs a way to respond to it. The paper and project page below give the experimental details and comparisons.