When to Switch: Reliable Action-Chunk Extension for Vision-Language-Action Models

When to Switch. Reliable Action-Chunk Extension for Vision-Language-Action Models. By Seonghoon Yu and six others, from KAIST, the Gwangju Institute of Science and Technology, and two other institutions. Published on the fifth of October. Its code repository is a placeholder: a readme, and nothing to run yet.

Vision-Language-Action models, or VLAs, serve as unified policies for robotic manipulation, but their inference is expensive. The robot has to pause between policy calls, and the result is stop-and-go execution: motion that keeps interrupting itself, and tasks that take longer to complete. The direct remedy is a longer action chunk. Predict more actions per call, and there are fewer calls and therefore fewer pauses. The catch is that predicting farther into the future makes long-chunk execution unreliable, so the time saved on pauses is paid back in failures.

Before proposing a fix, the authors analyze where that unreliability arises. Action errors inside long chunks are not spread evenly. They concentrate around the transitions between manipulation subskills, and they grow sharply as the chunk gets longer. That points at one quantity: transition timing, meaning when, within a chunk, to switch from one subskill to the next. RACE, short for Reliable Action-Chunk Extension, is built around that observation. It learns the transition timing, and it conditions action generation on what it has learned.

The timing is predicted from an auxiliary one-step denoising pass. That prediction is then used as a condition when the actions themselves are generated. The abstract describes the effect rather than the internals: by learning and conditioning on transition timing, RACE reduces errors at subskill transitions, and that is what allows longer action chunks to execute reliably. The model is told, in effect, where the hard moments in a chunk are going to fall.

Across simulation benchmarks, RACE outperforms fine-tuning at the same chunk length. With chunks twice as long, it surpasses recent state-of-the-art and efficient VLAs in success rate. With chunks four times as long, it remains competitive with them.

On a real robot, RACE runs with chunks four times longer. That reduces the idle time caused by stop-and-go execution by about a factor of five, while it still reaches a higher success rate than fine-tuning with the same chunk length. Code and a real-robot demo are public.

Read as a chain, the argument has four links. Pauses come from frequent policy calls. Longer chunks remove calls. Long chunks fail at subskill boundaries. Predicting when those boundaries fall makes the long chunks dependable again. Each step of the fix answers the failure exposed by the step before it, and the measurement that started it was simply looking at where, inside a chunk, the errors were.