PaperlayerView exact revision on arXiv

Source-linked AI summary

VLA-Rolling: Fewer Denoising Steps, Better Control via Adaptive Action Reuse

Sicheng He, Yi Pan, Qiuting Yu, Jiaming Huang, Qi Lu, Shuyan Wu, Hongyan Feng, Xuanyu Liu, Jun Zhang, Jie Zhang, Wenqi Zhang, Xuhong Zhang, Jianwei Yin

arXiv:2610.25043v3cs.AIcs.RO

TL;DR

Receding-horizon VLA execution discards the unexecuted action tail and restarts from Gaussian noise, although the tail may retain denoising progress and may be stale under updated observations. VLA-Rolling inherits aligned action tails, estimates per-action reuse confidence, and uses confidence to guide renoising and denoising-step allocation while keeping the VLA frozen. Across LIBERO, MetaWorld, LIBERO-Plus, and RoboTwin 2.0, VLA-Rolling reduces action-chunk inference time in all 16 pairs, achieving up to a 1.48× speedup (32.52% lower cost), and improves success in 14.

  • Problem

    Receding-horizon VLA execution discards the unexecuted action tail and restarts from Gaussian noise, although the tail may retain denoising progress and may be stale under updated observations.

    Source: §1 · pp. 1–2

  • Method

    VLA-Rolling inherits aligned action tails, estimates per-action reuse confidence, and uses confidence to guide renoising and denoising-step allocation while keeping the VLA frozen.

    Source: §1 · p. 2

  • Results

    Across LIBERO, MetaWorld, LIBERO-Plus, and RoboTwin 2.0, VLA-Rolling reduces action-chunk inference time in all 16 pairs, achieving up to a 1.48× speedup (32.52% lower cost), and improves success in 14.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Action tails can serve as reusable intermediate states for generating high-quality actions with fewer denoising steps.

    Source: §5 · p. 10

    The work has limited exploration of cross-environment generalization and calls for broader real-world evaluation across embodiments and interaction settings.

    Source: F Limitation and Future Work · p. 18

Abstract

from arXiv · show

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. We show that this discarded tail retains transferable denoising progress: action-tail initialization can match or outperform a native 10-step Gaussian baseline using fewer denoising steps. However, the optimal denoising budget varies substantially across samples, indicating that tail reuse should be adapted to the updated environment. Based on this observation, we introduce VLA-Rolling, a lightweight inference framework that reuses action tails through three stages: action-tail inheritance, per-action confidence estimation, and confidence-guided refinement through renoising and denoising budget allocation, while the underlying VLA remains frozen. Across $\pi_{0.5}$, SmolVLA, X-VLA, and MolmoAct2 on LIBERO, MetaWorld, LIBERO-Plus, and RoboTwin 2.0, VLA-Rolling reduces action-chunk inference cost in all 16 model--benchmark pairs, achieving up to a $1.48\times$ speedup, while improving success rate in 14 pairs by up to $15.08$ percentage points. These results show that action tails are not merely discarded predictions, but reusable intermediate states that enable more efficient and accurate VLA inference. These results demonstrate that the action tail can support the generation of better actions with fewer denoising steps. Our project is available at https://chenbao123ww.github.io/VLA-Rolling/.

1 Introduction

Receding-horizon VLA execution discards action tails that may retain useful denoising progress, but their usefulness depends on how the environment changes. VLA-Rolling adapts tail reuse with confidence-guided refinement, improving inference efficiency and often success across the evaluated model–benchmark pairs.

1 Introduction

  • The unexecuted action tail overlaps in time with actions needed in the next inference round and may provide a better starting point than fresh Gaussian noise.Because the tail was generated from an earlier observation, it may also be stale.Source: §1 · pp. 1–2
  • The paper asks whether action tails can generate new actions faster without sacrificing accuracy.Receding-horizon execution runs a short action prefix, observes again, and replans; the remaining tail is discarded.Source: §1 · pp. 1–2
  • Tail initialization matches or improves the 10-step Gaussian baseline in four denoising steps for π0.5 and six for SmolVLA, but optimal budgets span 1–10 steps.This variation makes reuse dependent on how much the tail should be trusted under each new observation.Source: §1 · p. 2
  • VLA-Rolling inherits aligned action tails, estimates per-action reuse confidence from updated observations, and uses confidence to guide renoising and denoising-step allocation while keeping the VLA frozen.The confidence estimator is trained offline on temporally adjacent demonstrations.Source: §1 · p. 2
  • Inference cost fell in all 16 model–benchmark pairs, with up to a 1.48× speedup, while success improved in 14 pairs by up to 15.08 percentage points.Ablations expose the risk of naive reuse, while real-world target changes trigger additional refinement.Source: §1 · pp. 2–3

2 Probing the Feasibility of Action Tail Reuse

Receding-horizon execution refreshes observations but leaves an unexecuted action tail that conventional inference discards. Tail initialization can match or improve a 10-step Gaussian baseline with fewer steps, but the optimal refinement budget varies by sample.

2 Probing the Feasibility of Action Tail Reuse

  • Receding-horizon execution refreshes observations after a short action prefix, addressing accumulated mismatch between the robot’s actual state and later predicted actions.The remaining L = H − K actions form the tail, which is conventionally discarded after observations are refreshed.Source: §2 · p. 3
  • The diagnostic compares tail and fresh-Gaussian initialization against the same target action, aligning the inherited tail with the demonstration at the updated state.The tail comes from a frozen VLA prediction in the earlier state after the first K actions are assumed executed.Source: §2 · p. 3
  • Four denoising steps for π0.5 and six for SmolVLA match or improve the 10-step Gaussian baseline on both benchmarks.The inherited tail can also become less accurate with additional refinement: for π0.5 on LIBERO, the error ratio changes from 1.073 at six steps to 1.037 at eight.Source: Figure 2 · p. 3, §2 · p. 4
  • Optimal denoising steps span 1–10; even above the medians of 8 for π0.5 and 9 for SmolVLA, 33.2% and 42.6% of samples need further refinement.The standard deviations are 2.28 for π0.5 and 2.88 for SmolVLA, motivating a sample-specific budget.Source: §2 · p. 4, Figure 3 · p. 4

3 VLA-Rolling: Adaptive Action Tail Reuse

VLA-Rolling reuses the unexecuted action tail across inference rounds while the underlying VLA remains frozen. It estimates per-action confidence to guide tail preservation, renoising, and the denoising budget.

3 VLA-Rolling: Adaptive Action Tail Reuse

  • VLA-Rolling has three stages: inherit the tail, estimate action confidence from state and environmental change, then refine through renoising and adaptive denoising.The framework retains the prior environment state alongside the inherited, temporally aligned tail.Source: §3 · p. 4, Figure 4 · p. 5, §3.1 · p. 5

3.2 Action Confidence Estimation

  • The Action Confidence Estimator predicts the reuse reliability of each inherited action from the current environmental state and its changes.Its output c represents the estimated reliability of retaining the inherited action at each position.Source: §3.2 · p. 5
  • Confidence targets derive from inherited-action error against aligned demonstrations: smaller error indicates easier refinement and should receive higher confidence.The training objective penalizes overestimated confidence more heavily than underestimated confidence.Source: §3.2 · pp. 5–6

3.3 Confidence-Guided Refinement

  • Confidence guides both renoising, which controls injected Gaussian noise, and adaptive-step denoising, which allocates the action-generation budget.High-confidence positions retain more temporal information with less noise; low-confidence positions are renoised more strongly, approaching native VLA initialization.Source: §3.3 · p. 6
  • The denoising budget is bounded by Nmin and Nfull, with Nfull equal to the frozen VLA’s default denoising steps.The allocation uses the minimum predicted confidence across action positions.Source: §3.3 · p. 6
  • Each inference round rolls the remaining tail forward, so actions are continuously refined over multiple rounds before execution.Table 1 compares average task success rate (SR) and inference time per action chunk (Cost).Source: §3.3 · p. 6, Table 1 · p. 7

4 Experiments

Across the reported benchmarks, VLA-Rolling reduces inference cost while maintaining or improving success; ablations, shift analyses, and a real-world mutation task examine its components and behavior.

4.1 Overall Performance

  • Inference cost falls by 5.00%–32.52% across models and benchmarks, with task success basically unchanged or slightly improved.The reported throughput increase reaches 1.48×.Source: §4.1 · p. 7

4.2 Ablation Study

  • Success-rate improvements occur only in variants with renoising; action-tail inheritance alone lowers success by 4.89%, and adjusting denoising steps alone lowers it by 4.29%.The authors note that high-step refinement can impair tail quality and that renoising can help when actions lie outside the training distribution.Source: §4.2 · p. 8, Table 2 · p. 8

4.3 Analysis Experiments

  • The confidence estimator remains correlated with environmental change until shifts exceed what its default K=10 setting can comprehend, indicating robustness with limitations.Confidence is negatively correlated with the drasticness of environmental changes.Source: §4.3 · p. 8, Table 3 · p. 8
  • Across MetaWorld difficulty levels, successful trajectories require more denoising steps than failed trajectories, by up to 2.33 steps.The authors interpret this difference as evidence that confidence predictions differentiate trajectory outcomes and relate to action-generation difficulty.Source: §4.3 · pp. 8–9, Figure 5 · p. 9

4.4 Real-World Experiments

  • In the real-world potato-placement task, a sudden target-color change produces a confidence decrease, after which the robot switches from placing into the blue plate to the brown plate.The change is still within the task seen during training.Source: §4.4 · pp. 9–10, Figure 6 · p. 9

5 Conclusion

The paper concludes that VLA-Rolling can reuse action tails to generate high-quality actions efficiently. It situates this approach amid concerns about discarded actions, indiscriminate chunk reuse, and efficiency methods that trade away accuracy.

5 Conclusion

  • VLA-Rolling uses a lightweight plugin to estimate action-tail confidence and select denoising steps, reusing tail motion information across inference rounds.The conclusion reports that it predicts the denoising steps required for action tails and generates high-quality actions efficiently.Source: §5 · p. 10

A.1 Continuous Action Generation

  • Because only a portion of a generated chunk is executed and the remainder is discarded, avoiding that computational waste is an optimization problem for VLA models.Flow-matching transfers vision-language-model priors to guide continuous future-action generation.Source: §A.1 · p. 13

A.2 Action Chunk Reuse

  • Prior action-chunk reuse can preserve temporal consistency, but reusing low-quality chunks risks incorrect generation, motivating selective use and action-level evaluation.The passage cites prior work for temporal consistency and action-level evaluation.Source: §A.2 · p. 13

A.3 Efficient VLA Inference

  • The paper frames a further efficiency gap: model-size approaches require training from scratch, while decoder-inference approaches typically sacrifice action accuracy.The cited passage says few works can be universally applied to existing models while improving performance.Source: §A.3 · p. 13

B Method Details

The method details describe a confidence predictor that operates alongside a frozen VLA, and adaptive denoising that treats inherited tails as intermediate flow states for refinement.

B.1 Details of Model Architecture

  • The confidence predictor is an approximately 30M-parameter action-centric Transformer, while the backbone VLA remains frozen.It uses visual and proprioceptive representations and their temporal differences, the inherited tail, and the last executed action; its confidence output guides noise and tail-preserving behavior.Source: §B.1 · p. 13

B.2 Details of Action Confidence Estimator

  • The estimator input is padded to chunk length H with Gaussian noise, then attention-masked so the added noise does not affect judgment of the action tail.The output is truncated to tail length L; the simulator settings use H=30, K=10, L=20 for X-VLA and H=50, K=10, L=40 for the other listed models.Source: §B.2 · p. 14

B.3 Details of Adaptive-Step Denoising

  • Adaptive denoising treats the inherited tail as a partially denoised intermediate state whose starting flow time is set by its remaining refinement budget.Refinement then follows the native VLA Euler update rule, preserving consistency with its timestep embedding and discretization.Source: §B.3 · p. 14

C.1 Training Protocal

The protocol specifies the confidence model's training configuration and the benchmarks, episode counts, and inference-cost components used for evaluation.

C.1 Training Protocal

  • Training uses batch size 64 for 4,000 steps, with AdamW learning rate and weight decay both 10^-4 and asymmetric confidence-loss weights w+=2 and w-=1.The reported runs used one NVIDIA RTX PRO 6000 Blackwell Server Edition GPU.Source: §C.1 · p. 14

C.2 Inference protocol

  • Evaluation covers LIBERO, MetaWorld, LIBERO-Plus, and RoboTwin 2.0, measuring average success rate and per-chunk inference time.Inference cost includes VLM KV-cache generation, action-expert denoising, and confidence prediction; the reduction in adaptive denoising steps lowers the denoising component.Source: §C.2 · p. 15

D.1 Ablation Sensitivity

The calibration ablation identifies α=1.5 as the strongest tested balance of success and denoising budget, while the step-conversion ablation favors using minimum confidence to allocate denoising steps.

D.1 Ablation Sensitivity

  • At α=0, success was 97.25% with 6.861 denoising steps and 206.7 ms/block; at α=1.0, it was 97.00% with 6.155 steps and 187.6 ms/block.At α=2.0, success was 97.25%, and success remained above 97% across the tested calibration range.Source: §D.1 · p. 15
  • At α=1.5, success reached 98.50% with an average of 6.332 denoising steps, the highest success among tested settings and the default used in major experiments.The authors characterize this setting as the best balance between action quality and inference efficiency.Source: §D.1 · p. 15

D.2 Ablation on Denoising Step Convert

  • Using minimum confidence as the basis for denoising-step allocation better met the goal of avoiding accuracy loss, so it was selected as the default conversion strategy.Source: §D.2 · p. 15

D.3 Ablation on action confidence estimator parameters

The ablations favor a moderate-sized confidence estimator and indicate that the predicted denoising budget tracks refinement needs; adjacent execution contexts also remain similar enough for selective tail reuse.

D.3 Ablation on action confidence estimator parameters

  • Increasing the confidence estimator from 10M to 30M parameters raised average success from 97.25% to 98.00%, while inference cost rose from 1.267 ms to 1.579 ms.Scaling to 100M parameters added inference cost without further performance improvement; 30M was the default configuration.Source: Table 7 · p. 16, §D.3 · p. 16

D.4 Ablation on N* Nearby Permutation

  • Changing the predicted denoising budget by −1 or +1 reduced success by 0.83 or 1.08 percentage points on LIBERO, and by 1.21 or 0.61 points on MetaWorld.The authors interpret this local sensitivity as evidence that the predicted budget is aligned with individual action-tail refinement demand.Source: Table 8 · p. 16, §D.4 · p. 16

D.5 Environmental similarity statistics

  • Adjacent contexts after 10 execution steps remain highly similar, supporting selective reuse of action-tail information rather than discarding it and restarting from Gaussian noise.Rolling still incorporates newly observed biases and helps prevent error accumulation.Source: Table 9 · p. 17, §D.5 · p. 17

E.1 Real-World Experiment Setting

The real-world evaluation tests whether VLA-Rolling can respond to a changed placement target, and its training protocol accounts for latency-dependent execution horizons.

E.1 Real-World Experiment Setting

  • The real-world task perturbs placement by replacing the blue plate with a brown one as the robot approaches the target, requiring it to continue placing the potato in the new plate.The evaluation uses a pi05-base backbone and an AgileX PiPER 6-DoF arm.Source: §E.1 · p. 17
  • The instruction omits plate color, and the authors collected 20 trajectories for each target color before fine-tuning pi05 and training the confidence estimator with the VLA frozen.Source: §E.1 · p. 17

E.2 Real-World Training Protocol

  • Because adaptive denoising makes inference latency variable, the protocol estimates latency from its components and derives execution horizons using control frequency.For general models, the estimated horizon distribution bounds do not exceed 100 ms; training pairs were constructed on these horizons, followed by one training epoch.Source: §E.2 · pp. 17–18

F Limitation and Future Work

The paper identifies cross-environment generalization as an area with limited exploration and calls for broader real-world evaluation.

F Limitation and Future Work

  • Cross-environment generalization has limited exploration; future real-robot work could test more complex tasks, additional embodiments, and interaction settings.The authors also identify combining VLA-Rolling with other inference-acceleration mechanisms as a direction for exploration.Source: F Limitation and Future Work · p. 18

Related papers

Loading 2610.25043v3…