Humanoids in difficult terrain: Marco Hutter and Vladlen Koltun make planned contacts with the environment the interface between perception and movement

A careful planner predicts where the robot will lay hands and feet, a controller translates this plan into whole body movement: a single policy, without a deployment path.

To make a humanoid move forward on rough terrain, two families of methods are opposed: reinforcement learning without movement data, which generalizes well but hardly produces natural coordination of the whole body, and reference movement tracking, which is natural but unable to go beyond the trajectories seen in training. A team including Marco Hutter (ETH Zurich) and Vladlen Koltun proposes to combine the advantages of both by using movement data only as a priori during training.

The system is based on a self-supervised contact planner, based on an attention mechanism, which predicts the next points of contact between the body and the environment, and on a controller that transforms this plane into coordinated movement. The contact thus becomes the interface between what the robot perceives and what it does: politics learns the structure of locomotion rather than copying kinematics. A refinement step then makes it possible to explore environmental variations that are absent from the original data. When deployed, the robot uses only its on-board sensors, without any movement references.