The training pipeline uses TCRS supervision, an offline synthesizer that combines raw motion clips with sampled height fields to create terrain-conformal references. At deployment, an identity-gated vision student preserves the raw command and adds terrain corrections through zero-initialized residual pathways only where the terrain requires them, avoiding per-skill tuning at test time.
Perceptive BFM is useful for humanoid locomotion, motion imitation, terrain adaptation, fall recovery, and robot demonstrations. The project includes a browser demo powered by MuJoCo WASM and ONNX Runtime plus a Unitree G1 gallery covering backflips, cartwheels, stair walking, expressive motion, obstacles, outdoor surfaces, and recovery behaviors.
