Event
GRAFT: Terrain Residual Adaptation of Frozen Behavior Trackers
1 reports · 1 sources
Liang Li and co-authors propose the GRAFT framework, which, without modifying the parameters of a frozen humanoid behavior foundation model (BFM) or the motion reference interface, uses a lightweight adapter to apply terrain-conditioned corrections only to the lower body's execution targets. The adapter is trained with task and foot geometry rewards and is constrained to produce no superfluous corrections on flat ground; it also distills from privileged terrain observations to onboard depth. The authors state that they instantiate the framework on four frozen trackers and evaluate it in simulation on stair, trench, and platform scenarios; they also conduct real-robot validation on a Unitree G1, where the adapted SONIC tracker completes terrain traversal relying on onboard depth and unmodified references.
Report timeline
- GRAFT: Geometry-Aware Residual Adaptation for Frozen Humanoid Behavior Trackers
GRAFT proposes a geometry-aware residual adaptation framework that, without changing the parameters of a frozen humanoid behavior foundation model (BFM) or its motion-reference interface, uses a lightweight adapter to apply terrain-conditioned corrections only to lower-body execution targets. The adapter is trained with task and foot-geometry rewards, is regularized so it produces no unnecessary corrections on flat ground, and is distilled from privileged terrain observations to onboard depth.
arXiv cs.RO ·