When to Switch: Reliable Action-Chunk Extension for Vision-Language-Action Models
RACE targets the moment a robot should switch subskills so longer action chunks do not break at transitions.
The paper says VLA robots lose time because expensive policy calls create stop-and-go motion. Longer action chunks reduce those pauses, but the authors find errors pile up around subskill transitions and worsen as chunks get longer. RACE predicts transition timing with an auxiliary one-step denoising pass, then conditions action generation on that timing. In simulation it beats same-length fine-tuning, and on a real robot it uses 4x longer chunks while cutting idle time from stop-and-go execution by about 5x. HF Daily Papers' note
The paper says VLA robots lose time because expensive policy calls create stop-and-go motion. Longer action chunks reduce those pauses, but the authors find errors pile up around subskill transitions and worsen as chunks get longer. RACE predicts transition timing with an auxiliary one-step denoising pass, then conditions action generation on that timing. In simulation it beats same-length fine-tuning, and on a real robot it uses 4x longer chunks while cutting idle time from stop-and-go execution by about 5x. HF Daily Papers' note
score 4