When to Switch: Reliable Action-Chunk Extension for Vision-Language-Action Models
RACE predicts subskill transition timing so VLA models can execute longer action chunks with fewer pauses.
Vision-language-action policies pause between expensive inference calls, and simply lengthening action chunks becomes unreliable near subskill transitions. RACE predicts transition timing with an auxiliary one-step denoising pass and conditions action generation on it. With 2x longer chunks it beats recent efficient VLAs in simulation success rate, and with 4x chunks it stays competitive. On a real robot, 4x chunks cut stop-and-go idle time about 5x while exceeding same-length fine-tuning success.
- Long-chunk errors concentrate at transitions between manipulation subskills.
- RACE predicts transition timing from an auxiliary one-step denoising pass.
- Two-times longer chunks beat recent efficient VLAs in simulation success rate.
- Four-times chunks cut real-robot idle time about fivefold versus stop-and-go execution.
From the original article
1 min read at huggingface.co
Vision-Language-Action (VLA) models serve as unified policies for robotic manipulation, yet their expensive inference forces robots to pause between policy calls, resulting in stop-and-go execution that interrupts smooth motion and prolongs task completion. Extending the action chunk reduces policy calls and hence these pauses, but predicting farther into the future makes long-chunk execution unreliable.
By Seonghoon Yu, Dongwon Kim, HyungRok Jung · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.