Event

VLA-Rolling reuses action tails to reduce denoising steps, achieving up to 1.48× speedup across 16 experiment settings

1 reports · 1 sources

Sicheng He et al. propose the VLA-Rolling inference framework, which, without modifying the underlying VLA model (keeping it frozen), reuses action tails through three stages—action tail inheritance, per-action confidence estimation, and confidence-guided refinement based on re-noising and denoising step allocation—to reduce the number of denoising steps in action chunk inference. Across 16 combinations of four models (π0.5, SmolVLA, X-VLA, MolmoAct2) and four benchmarks (LIBERO, MetaWorld, LIBERO-Plus, RoboTwin 2.0), the framework reduced inference cost in all cases, with a maximum speedup of 1.48×; in 14 of these settings, the success rate improved, by up to 15.08 percentage points. The authors state that action tail initialization can match or exceed the effect of the native 10-step Gaussian baseline with fewer denoising steps.