Behavioral foundation models (BFMs) aim to give humanoids a reusable repertoire of whole-body behaviors. But in real-world control, the same behavioral intent may require fundamentally different motions under different terrain conditions. A new preprint introduces P-BFM (Perceptive Behavioral Foundation Models), which uses unsupervised reinforcement learning to incorporate perception into the behavior selection process.
The paper argues that a perceptive model can choose or generate motions that fit the current ground conditions rather than relying on a fixed behavior set. This is a single arXiv preprint, so the claims are not yet peer-reviewed, and no independent sources are available to compare or contrast with the authors' approach.