Paper review · October 22, 2026
VLA-Rolling: Reusing discarded action tails to control robots with fewer denoising steps
VLA-Rolling: Fewer Denoising Steps, Better Control via Adaptive Action Reuse
VLA-Rolling reuses the later part of the actions that flow-matching VLAs discard without executing, using it as the starting point for the next generation according to its reliability. This reduces inference cost while also raising success rates in most cases.
Problem
Flow-matching VLAs build an action chunk by denoising noise over multiple steps, but execute only the first part before replanning with a new observation. The unexecuted later part (the action tail) is discarded, and the next chunk is generated again from fresh Gaussian noise. The denoising computation already completed is effectively wasted.
Method
In diagnostic experiments, starting from the action tail let π0.5 match the action error of the 10-step Gaussian baseline in 4 steps, and SmolVLA in 6 steps. However, the optimal number of steps per sample varied widely, from 1 to 10. Based on this, VLA-Rolling works in three stages. (1) It carries over the previous tail, aligned to the next planning window. (2) A lightweight reliability estimator looks at the new observation and state change and predicts a per-action reuse reliability. (3) It mixes more noise into low-reliability actions and sets the number of denoising steps according to the lowest reliability in the chunk. The estimator is trained offline on temporally adjacent demonstration data and is penalized more heavily for overestimating reliability. The VLA itself is kept frozen.
Results
- Across all 16 combinations of π0.5, SmolVLA, X-VLA and MolmoAct2 with LIBERO, MetaWorld, LIBERO-Plus and RoboTwin 2.0, inference time per action chunk fell by 5.00–32.52% (up to 1.48x throughput).
- Success rates rose in 14 of the 16 combinations, with the largest gain on SmolVLA LIBERO-Plus, from 20.74% to 35.82% (+15.08 percentage points).
- On π0.5 LIBERO, the success rate went from 96.38% to 97.83%, and inference time from 168.09 ms to 129.60 ms.
- In ablations, simply carrying over the tail lowered the success rate by 4.89%, and adjusting only the number of steps lowered it by 4.29%. Improvements appeared only in variants that included renoising.
Limitations
Success rates dropped slightly for X-VLA on LIBERO-Plus (-0.13 percentage points) and RoboTwin (-0.28 percentage points). In some combinations, the cost savings are only 5–8%. The reliability estimator must be trained separately on demonstration data. Evaluation is mainly in simulation, and real-robot experiments could not be examined in detail within the portion of the paper reviewed. The change value reported for π0.5 MetaWorld in Table 1 does not match the success-rate difference, so the original paper should be checked.
Why it matters
This is a plug-in method that improves both speed and success rate by changing only the inference stage, without retraining the VLA. It is worth trying directly when denoising cost is the bottleneck in real-time robot control.
Key terms
- VLA (Vision-Language-Action) model
- A model that takes camera images and language instructions as input and directly outputs robot actions.
- Action chunk / Action tail
- An action chunk is a set of future actions predicted at once. The later part of it that remains unexecuted is the action tail.
- Receding-horizon execution
- An approach that reduces error accumulation by repeatedly executing only the first part of a long plan and then replanning with a new observation.
- Renoising
- A step that mixes noise into actions to be reused, in an amount inversely proportional to their reliability, giving room to correct outdated predictions.
Original
Authors: Sicheng He, Yi Pan, Qiuting Yu, Jiaming Huang, Qi Lu, Shuyan Wu, Hongyan Feng, Xuanyu Liu, et al. · Posted 2026-10-18 · License CC BY 4.0 · Based on the full text
This review is an AI-written summary of the paper. Please treat the original paper as the reference.