RACE Framework Lets VLA Policies Run 4x Longer Action Chunks, Cuts Robot Idle Time ~5x

RACE, posted October 6, tackles the stop-and-go problem in VLA models like π₀ and GR00T by predicting where subskill transitions fall inside an action chunk, then conditioning the generation step on those transitions via adaptive-RMSNorm gating. On a real robot with 4x longer chunks it hits 66% success vs 48% for the fine-tuned baseline and drops idle time ~5x; training is lightweight — 17 hours on 4 RTX A6000s, versus 1,828 hours for a comparable RL approach.