terrain-aware behavioral foundation model for humanoid whole-body controlExpert
P-BFM: Perceptive Behavioral Foundation Models for Humanoid Control via Unsupervised Reinforcement Learning
A team from RoboParty's lab and two Shanghai and Beijing universities trained one frozen humanoid controller that walks, tracks motions and recovers from pushes on stairs, slopes and boxes without terrain-specific demonstrations. In this preprint it recovered from 61.4% of perturbations on unseen motions, against 16.6% for the best baseline, and ran zero-shot on a Unitree G1 using its onboard depth camera.