PaceVLA Varies Action Timing to Preserve Precision

A vision-language-action policy can accelerate execution by using coarse actions when they are suitable and switching to fine-grained actions when the context requires precise control. PaceVLA, a framework described in a preprint by Hojin Jeon and co-authors, makes that choice at each inference step instead of applying one temporal resolution throughout a task.

Temporal action quantization combines consecutive low-level actions into coarser actions, reducing redundant steps during phases that do not need detailed motion. Applying it indiscriminately can discard important details during precision-sensitive interactions. PaceVLA therefore uses mode-specific action decoders, built on a shared action expert, to support multiple temporal granularities and prediction horizons. A quantizability signal determines whether coarse execution is appropriate, while a context-conditioned router selects the execution mode.

The authors report that PaceVLA improved execution efficiency while maintaining task success rates across three multi-task simulation benchmarks and six physical-robot tasks. Their evaluations covered multiple robot embodiments as well as long-horizon and dexterous manipulation. The available abstract does not provide numerical gains or detailed comparisons, and these results remain claims from the preprint authors rather than independent confirmation.