Event

Proposes terrain-aware humanoid control framework Perceptive BFM

1 reports · 1 sources

Zifan Wang et al. proposed Perceptive BFM on arXiv, a terrain-aware whole-body control framework for humanoid robots that adapts human motion priors to terrain observations from the robot's perspective. The framework retains the original kinematic reference as the behavioral interface and uses local terrain observations to adapt contact, posture, and timing. In terms of implementation, the student model is an identity-gated Transformer tracker, with terrain features entering through a residual pathway, initialized to preserve the original motion-tracking prior and to produce local corrections only when needed; during training, a teacher that blindly adapts the reference is trained first, and then terrain-consistent behaviors are transferred to the original-reference student model used for deployment through action alignment in the target coordinate frame.