Humanoid Mobility via Attention-based Contacts

Can one humanoid policy combine the natural whole-body motion of reference-tracking controllers with the terrain generality of reinforcement learning trained without references?

Pure reinforcement learning generalises across terrain but rarely discovers coordinated whole-body moves such as using hands on a ladder, while controllers that track recorded human motion look natural but fail once the terrain differs from the clips they learned.

Earlier systems either shaped rewards by hand for each skill, or tracked full reference trajectories of root and joint motion, which ties the robot to the recorded motion. Motion-generation planners such as MGMT predict future hand and foot positions but struggled on terrain without matching data.

The team uses motion data only as a training scaffold. Stage 1 pretrains a whole-body tracker that follows contact commands (where each hand and foot should touch down) labelled from retargeted motion clips, with a backward curriculum along each trajectory. Stage 2 trains an attention-based planner that picks those contact targets from a terrain map. Stage 3 fine-tunes planner and tracker together with reinforcement learning on goal-reaching tasks, including terrain with no reference motion at all.

On terrain covered by the motion data, contact commands alone gave at least 98.5% success in every category, with no gain from also feeding reference joint motion. Stage 3 fine-tuning raised success and contact accuracy in every goal-reaching and unseen category, by 54.8 percentage points on sparse whole-body terrain. On eight composite courses the method averaged 92.5% success against 58.8% for MGMT, over 1,000 episodes per method and course. The attention planner beat an MLP planner after fine-tuning, most clearly on unseen combinations and distractor objects. On the real G1 the same policy climbed a ladder and a stage, vaulted, crawled, jumped down and sat down, with perception, mapping and inference onboard.

This is a preprint. The success figures are simulation results with sensor noise, physics randomisation and action delay switched off; the hardware tests are demonstrations rather than measured success rates. The authors say the terrain and contact representation is 2.5-D, so overhead contacts, hanging and manipulation need richer 3D perception, the system only reaches local goals rather than navigating globally, and it cannot handle aerial moves without distinct contacts.