P-BFM: Perceptive Behavioral Foundation Models for Humanoid Control via Unsupervised Reinforcement Learning

P-BFM is a behavioral foundation model that enables humanoids to adapt their motions to different terrains by separating global transport from behavioral intent. It uses unsupervised reinforcement learning to generate terrain-appropriate behaviors without terrain-specific demonstrations, improving motion fidelity and recovery across challenging terrains. The model supports zero-shot transfer to new platforms using onboard depth sensing.