acceptodds
Under review as a conference paper at ICLR 2027

Accelerating VLA Inference with Perturbation Informed Execution

Abstract

Action-chunking vision-language-action (VLA) models predict sequences of future robot actions. Under fixed-horizon execution, the robot executes a fixed-length prefix of each predicted chunk before invoking the policy again, leaving the remaining predictions unused. For a given number of executed actions, short execution horizons require more policy calls. Extending the execution horizon amortizes policy inference cost over more actions. However, uniformly applying long horizons across episodes can limit opportunities for precise corrections. To address this limitation, we propose PICES (Perturbation-Informed Chunk and Execution Horizon Selection), a training-free method that jointly selects an action chunk and the number of actions executed before the next policy call. Within a single policy call, the action expert generates a batch of action chunks from multiple initial noise samples and their locally perturbed counterparts, using the same conditioning features across all samples. PICES exploits differences between corresponding actions in each nominal chunk and its counterpart to determine candidate-specific execution horizons and select a nominal chunk for execution. Experiments on simulation benchmarks and real-world manipulation tasks demonstrate that PICES achieves speedup by up to 2.3 in wall-clock policy inference time on Jetson Orin and Thor relative to fixed-horizon baselines while improving the success rate. The code is available at : https://anonymous.4open.science/r/PICES_VLA-B1BF

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action, Accelerating VLA, Dynamic Action Chunking

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_dd65OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces PICES, a training-free inference method for flow-matching VLA policies that generates multiple candidate action chunks and their perturbed counterparts in a single batched policy inference. It measures action sensitivity to perturbations and uses the scores to jointly select the action chunk and the execution horizon before replanning. The method exploits the memory-bound nature of action experts on Jetson hardware, where the lowered total inference time by reducing the number of policy calls outweighs the relatively little overhead incurred by generating additional candidates. Experiments on simulations and a real-world robot report fewer policy calls with comparable or slightly improved success rates relative to fixed-horizon execution, achieving 1.2–2.3×1.2–2.3\times inference-time speedups.

Critical strengths
  1. The paper backs up its main idea: spending a few extra samples per call to make fewer calls. It shows that choosing well among multiple sampled chunks allows longer execution. It measures on the Jetson hardware that the extra samples add only a small delay, a cost already included in its reported speedups.
  2. On SimplerEnv and on the real robot, PICES matches the success rate of the best fixed horizon with fewer policy calls. The N=1N = 1 ablation shows that the sensitivity-based horizon rule works very well. With a single candidate, it already accounts for most of the reduction in calls.
Critical weaknesses
  1. PICES does not remove the need to tune the execution horizon. On LIBERO with π0.5\pi_{0.5}, a fixed horizon of 8 matches or outperforms the default PICES. On LIBERO and on the real robot, PICES outperforms the best fixed horizon in both success and speed only when its minimum horizon is set to 8, which is the best fixed horizon in the same evaluations.
  2. The success gains over fixed horizons are small, such as 68.7 against 68.5 on SimplerEnv and 93.5 against 90.7 on the real robot, and no error bars or repeated runs are reported. The claim of improving the success rate should read as matching it.
Questions for authors
  1. On the real robot, W=8W = 8 is the best fixed horizon from the sweep, and WW is also the shortest horizon PICES can choose. What does PICES reach with a smaller WW, so that the rule has to find the right horizon on its own?
  2. Table 1 averages policy calls over successful episodes only, so methods with different success rates are averaged over different episodes. How do call counts and speedups change when including failed episodes as well?
Reviewer_6hCBOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces PICES, a training-free method that jointly selects an action chunk and its execution horizon for VLA policies. It perturbs the initial noise and compares the resulting action predictions to estimate per-step sensitivity, which serves as a proxy for action reliability. PICES filters candidates with low initial sensitivity, estimates their execution horizons, and selects the one with the longest horizon to reduce policy calls and cumulative inference time. The authors evaluate the method across multiple VLA models on SimplerEnv, LIBERO, and real-world UR10e manipulation tasks. They also examine the effect of candidate count and compatibility with visual token compression. The reported results show improved average success rates and lower cumulative inference time compared with the evaluated fixed-horizon baselines.

Critical strengths
  1. The method is simple and requires no additional training. PICES compares action predictions before and after perturbing the initial noise to estimate how long each candidate can be executed, then selects the candidate with the longest estimated horizon. This connects action selection with execution length and makes the method straightforward to apply to existing pretrained policies.

  2. The efficiency evaluation accounts for the method’s own overhead. The authors measure not only policy calls but also the cost of candidate generation, sensitivity computation, and selection. Across multiple models, simulation benchmarks, and real-world tasks, PICES achieves higher average success rates and lower cumulative inference time than the fixed-horizon baselines, providing empirical support for its efficiency gains.

Critical weaknesses
  1. It is still unclear how much of the improvement comes from the proposed sensitivity metric. PICES differs from AAC, AutoHorizon, and A3 not only in its reliability signal, but also in how it selects action chunks and determines execution horizons. The comparisons demonstrate the performance of the full method, but do not isolate whether the sensitivity metric itself is more effective. Since this is a central contribution, I would like to see a controlled comparison that isolates its value relative to other reliability signals. The use of the initial-window mean sensitivity as the stopping threshold also needs more direct validation. The open-loop results offer some support, but do not clearly establish why this rule works or when it fails.

  2. The real-world experiments compare only against fixed-horizon baselines, without comparisons to existing adaptive methods. Table 3 evaluates several fixed execution horizons but does not include AAC, AutoHorizon, or A3. These results support the benefits of PICES over the tested fixed-horizon settings, but leave its advantage over existing adaptive approaches unclear. Since adaptive execution is already an established approach, this comparison is important for assessing what PICES adds in real-world deployment.

Questions for authors
  1. Appendix B.2 states that the timing includes sensitivity computation and candidate selection. Could you report the time spent on these two steps separately? I would like to understand whether the extra cost mainly comes from generating multiple action chunks or from evaluating and selecting them.

  2. The comparisons with AAC, AutoHorizon, and A3 differ not only in the reliability signal, but also in how candidates and execution lengths are selected. It is therefore unclear how much of the gain comes from the sensitivity metric itself. Could you compare different signals while keeping the other conditions as similar as possible?

  3. Why were AAC, AutoHorizon, and A3 not included in the real-world experiments? The current results show that PICES outperforms the tested fixed-horizon settings, but I would also like to know how it compares with existing adaptive methods. Are any such results available?

  4. Table 5 shows that PICES brings further gains when combined with FastV and GridS. Would AAC, AutoHorizon, or A3 provide similar gains? A comparison under the same compression settings would help clarify whether this benefit is specific to PICES or is generally shared by adaptive execution methods.

  5. Why is the mean sensitivity over the first WW steps used as the stopping threshold? Does exceeding this mean at a later step indicate that execution should stop? Could you provide a direct analysis linking threshold crossings to action error or failure risk?

  6. Table 1 reports average policy calls only for successful episodes, but success rates differ considerably across methods. Do the efficiency conclusions still hold when the cost of failed episodes is included? Could you also report total inference time across all episodes and divide that total by the number of successes to give a more complete comparison?

Reviewer_KpF9OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes PICES, a training-free method for accelerating action-chunked VLA policies by jointly selecting action chunks and their execution horizons. It estimates action reliability through sensitivity to perturbations in the initial noise and uses this information to select candidates that support longer execution horizons. Experiments across multiple VLA models, simulation benchmarks, and real-world manipulation tasks demonstrate improvements in inference efficiency while maintaining competitive task success rates.

Critical strengths
  • The paper identifies a meaningful trade-off between inference overhead and execution reliability in action-chunked policies. The motivation is clear, and the proposed method directly addresses the identified problem.

  • The method is intuitive, training-free, and relatively easy to integrate into existing VLA policies. The joint selection of action chunks and execution horizons is a reasonable design, supported by the motivational analysis.

  • The experiments cover multiple VLA backbones, simulation benchmarks, and real-world robot tasks. The hardware-level latency analysis on Jetson devices is particularly useful. Overall, the results provide convincing evidence that the method can reduce inference overhead under the evaluated settings.

Critical weaknesses
  • The main motivation assumes that policy inference introduces waiting time during robot execution. However, existing approaches such as Real-Time Chunking (RTC) and subsequent asynchronous execution methods can largely overlap policy inference with action execution, substantially reducing or eliminating this waiting time. This weakens the practical significance of reducing the number of policy calls. Although PICES may still reduce computational demand, its benefits for actual task execution remain unclear under asynchronous pipelines. This is particularly important because the reported speedups are based on cumulative inference time rather than measured end-to-end task completion time.

  • The motivational analysis assumes that actions closer to the ground-truth demonstrations are more reliable. However, robot manipulation is inherently multimodal, and multiple distinct action sequences can successfully accomplish the same task. A larger RMSE from a single reference trajectory does not necessarily indicate worse actions, nor does a smaller RMSE guarantee better task performance. Therefore, the RMSE-based analysis provides limited evidence for the claimed reliability of longer execution horizons.

  • The method assumes that actions with low sensitivity to initial-noise perturbations are more suitable for longer execution. However, local stability does not necessarily imply action accuracy or reliability: a policy can produce consistently incorrect actions. While the empirical results support the usefulness of this heuristic, the relationship between sensitivity and execution reliability could be more thoroughly investigated.

Questions for authors

Could the authors provide a more systematic analysis of the selected horizons across different manipulation phases, particularly free-space motion versus contact-rich interactions, rather than relying primarily on qualitative examples?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.