VLA-Rolling reuses action tails to cut VLA denoising steps
Original title: VLA-Rolling: Fewer Denoising Steps, Better Control via Adaptive Action Reuse
Across 16 frozen model-benchmark pairs it lowered chunk inference cost, with up to 1.48 times speedup and higher success in 14.
Flow-matching vision-language-action models denoise an action chunk but execute only a short prefix, discarding the rest. VLA-Rolling inherits that tail, estimates per-action confidence, and allocates further renoising only where needed, leaving the VLA frozen. On pi0.5, SmolVLA, X-VLA, and MolmoAct2 across LIBERO, MetaWorld, LIBERO-Plus, and RoboTwin 2.0, the paper reports lower inference cost in every pair.
- Tail initialization can match a native 10-step Gaussian baseline with fewer steps.
- The best denoising budget varies substantially from sample to sample.
- Peak reported speedup is 1.48 times; peak success gain is 15.08 points.
- Underlying VLA weights stay frozen during inference.
From the original article
1 min read at huggingface.co
Flow-matching vision-language-action (VLA) models generate an action chunk by iteratively denoising Gaussian noise, but typically execute only a short prefix before replanning with updated observations. Under this receding-horizon paradigm, the unexecuted remainder---the action tail---is discarded, and the next chunk is regenerated from fresh noise.
By Sicheng He, Yi Pan, Qiuting Yu · 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.