Humanoid behavior foundation models typically learn reusable whole-body control policies from broad human motion priors, assuming the reference motion is already physically compatible with the robot's surroundings. That assumption breaks when a human demonstrator, an operator, and a robot inhabit different environments: the human motion may specify the intended behavior, but not the footholds, clearance, body height, or contact timing required by the robot's local terrain.
To address this, the authors introduce Perceptive Behavior Foundation Model (Perceptive BFM), a terrain-aware control framework that keeps raw kinematic motion references as the interface while using local terrain observations to adapt contacts, posture, and timing. They develop terrain-conformal reference synthesis (TCRS), which converts locomotion-oriented human motion clips into terrain-consistent references through contact-aware foothold construction, foot-geometry-aware swing optimization, support-aware root reconstruction, collision repair, and multi-point inverse kinematics.
The training pipeline uses a blind adapted-reference teacher, then transfers terrain-conformal behavior to a deployed raw-reference student via target-frame action alignment. The student is an identity-gated Transformer tracker whose terrain features enter through residual pathways initialized to preserve the motion-tracking prior, so the model produces local corrections only when needed. This design is a step toward humanoid controllers that can reuse diverse human motion priors in real, uneven environments rather than assuming flat, compatible ground.