acceptodds
Withdrawn from ICLR 2027

Fast Thinking, Steady Control: Coherent Action-Chunk Transitions for Real-Time VLA Inference

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Vision-Language-Action (VLA) models have achieved impressive performance on diverse embodied tasks. To improve responsiveness to the environment, a growing body of work has focused on accelerating VLA inference. However, in real-world deployment, faster VLA inference does not necessarily translate into better control. Under continuous asynchronous inference, lower inference latency increases the frequency of chunk switching, which introduces inter-chunk discontinuities and latency-alignment errors, leading to discontinuous, jerky motion and even degraded task success. In this paper, we propose F2S-VLA, a training-free execution framework that translates fast asynchronous VLA inference into steady and coherent robot control. To mitigate inter-chunk discontinuity, we propose Noise-Decoupled Chunk Prediction, which generates two action chunks in a single forward pass with reference and newly sampled noise to separate observation-driven trajectory updates from noise-induced variations, and Anchor-Guided Trajectory Fusion, which uses the ongoing trajectory as an anchor to selectively fuse the new predictions, adaptively balancing responsiveness and continuity without task-specific thresholds. To eliminate latency-alignment errors, we introduce Latency-Aligned Action Resampling, which registers action chunks on a shared physical timeline and resamples them at the measured takeover time using the controller's native interpolation rule, improving temporal alignment granularity from tens of milliseconds to less than two milliseconds with negligible overhead. Experiments demonstrate that F2S-VLA outperforms RTC by 9.0% on RoboCasa365 benchmark and achieves a 93.9% average success rate across three long-horizon real-robot tasks. Our code is available on Anonymous Github.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action Models, Asynchronous Inference, Action Chunking, Real-Time Robot Control

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_dBY1OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

F2S-VLA addresses discontinuities that arise when fast asynchronous inference causes frequent action-chunk transitions. It generates predictions with reused and fresh noise, fuses them against a slowly refreshed anchor, and resamples actions on a shared physical timeline. The method requires no policy retraining.

Critical strengths
  • The paper identifies an important deployment issue that inference-speed studies can miss: faster predictions do not automatically produce better robot control. The analysis of chunk-switching frequency, stochastic variation, and fractional timing offsets is useful and directly relevant to asynchronous execution.

  • The reported improvements are substantial. Average RoboCasa365 success rises from 22.3% with fixed-horizon RTC to 31.3%, and the physical-robot results show gains across all three tasks. Component ablations also suggest that the anchor and resampling are consequential rather than cosmetic additions.

Critical weaknesses
  • The proposed observation/noise attribution is not fully justified. Writing N=P+(R−P)+(N−R)N=P+(R-P)+(N-R) establishes an identity, but interpreting the terms requires a precise relationship between the anchor, reference noise, and fused execution trajectory. This relationship needs clearer explanation. The categorical statement that anchor reuse “does not collapse action diversity” also reads as defensive reassurance; diversity is not directly measured.

  • Experimental reporting is incomplete. Tables 3–4 omit task/trial counts, seed coverage, and uncertainty estimates, making the reliability of the reported gains difficult to assess. The main success comparison also centers on π0.5, so the evidence for generality across VLA policies remains limited.

Questions for authors
  • How exactly are the anchor trajectory and its associated noise updated after fusion, and when is R−PR-P a valid observation-driven comparison?

  • What evaluation counts and uncertainty intervals support Tables 3–4, and do the gains hold on another VLA backbone?

Reviewer_bBgDOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper focuses on the topic of asynchronous execution paradigm for VLAs. The authors claim that faster inference may not be beneficial for better control, and identify two issues within asynchronous execution: inter-chunk discontinuity and latent-alignment errors. Motivated by these two challenges, the paper proposes F2S-VLA, a training-free framework that make consecutive chunks connect coherently. Experiments are conducted on the Robocasa365 simulation benchmark and real-world scenarios.

Critical strengths

The paper studies the important problem of asynchronous inference for VLAs, specifically in the aspect in making consecutive chunks more coherent. The proposed training-free method seems to be able to be integrated to different models. The experiment results show that the proposed method is able to outperform the selected baselines.

Critical weaknesses

I have serious concerns about the claims, evaluations, and conclusions presented in the paper.

  1. The paper is ambiguous about certain concepts and some claims are misleading. Specifically, the authors claim that "inter-chunk discontinuities, which become more evident under fast inference even with continuity-enhancing methods such as RTC" (line 79). This actually only happens for continuous asynchronous inference, which switches the executed chunk on every inference completion. For fixed-horizon asynchronous inference, the claim is not true. Fast inference will not lead to more frequent mode switching. More importantly, the term of "continuous asynchronous inference" has never been explained throughout the paper. Also, in disagreement with line 152, this continuous async paradigm is not conventional. Therefore, the lack of explanation and ambiguous presentation can make the paper overclaim its findings/statements, misleading readers and future research.

  2. For the first challenge the authors claim to identify: inter-chunk discontinuity, it is not something new even in terminology [1]. This issue persist in chunked policies instead only in asynchronous settings. The paper proposes to address this, but does not compare with important related works such as temporal ensembling [2] or BID [3], which are all test-time methods (refer to RTC codebase for their implementation).

  3. The second challenge, latency-alignment error, is an implementation-specific thing that does not necessarily has to exist. It can be simply solved by preserving a buffer at the robot side to store the finished and transmitted policy prediction results. In this way, inference completion and action takeover need not coincide. The controller can continue to execute the previous chunk and switch at an aligned control tick once the new chunk is available. If observations and action timestamps share the same time grid, the appropriate new-chunk index is determined exactly by the integer number of elapsed control steps, without truncating wall-clock inference latency. Moreover, the paper states that "The controller usually interpolates between consecutive waypoints" and "This error occurs at each chunk transition and accumulates over time", I do not think they are generally true. An execution stack may stream individual position targets, hold commands between updates, or send velocity or torque commands. Thus interpolation is not always existent and error accumulation is not always the case.

  4. The real-world experiments are very briefly mentioned. No image demonstrations, videos, or other types of evidence is provided. The experiment details are also insufficient. No robot type, inference hardware, actual delay measurement, and policy model is described.

  5. For all simulation evaluations, the d=0d=0 results should be included.

[1] Real-Time Robot Execution with Masked Action Chunking.

[2] Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware.

[3] Bidirectional Decoding: Improving Action Chunking via Guided Test-Time Sampling.

Questions for authors

Please refer to the above critical weaknesses. Beside them, I have the following questions:

  1. Line 70: "However, in real-world deployment, faster inference does not necessarily translate into better control performance. As shown in Fig. 1". Yet, Figure 1 shows simulation results, it is not real-world deployment.

  2. The details in how the simulation benchmarks are reconstructed as asynchronous execution benchmarks are unknown. It would be beneficial to clarify them.

Though given my current decision as clear rejection, I would consider raising the score if the authors could resolve my concerns.

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

This paper addresses inter-chunk discontinuities in asynchronous embodied models by proposing Noise-Decoupled Chunk Prediction and Anchor-Guided Trajectory Fusion to improve inter-chunk continuity, along with Latency-Aligned Action Resampling to eliminate latency-alignment errors. The effectiveness of the proposed method is evaluated on RoboCasa365.

Critical strengths
  • The paper proposes Noise-Decoupled Chunk Prediction and Anchor-Guided Trajectory Fusion to improve inter-chunk continuity, as well as Latency-Aligned Action Resampling to eliminate latency-alignment errors. These methods are evaluated on RoboCasa365.
  • The proposed method achieves higher success rates than RTC and VLASH on RoboCasa365. Its effectiveness is also validated through real-robot experiments.
  • The paper includes ablation studies to evaluate the effectiveness of the proposed components.
Critical weaknesses
  • Some statements in the paper are confusing. For example, the authors state that "lower inference latency increases the frequency of chunk switching, which introduces inter-chunk discontinuities and latency-alignment errors." However, lower inference latency does not necessarily lead to a higher chunk-switching frequency. Even if inference latency approaches zero, this does not imply that chunk switching occurs at an infinitely high frequency. This statement only holds under an additional assumption: a new chunk is adopted immediately after inference completes, followed by the immediate launch of the next asynchronous inference.
  • Table 3 and Table 4 only report success rates, while overlooking other important metrics for asynchronous inference, such as trajectory jerk and task completion time.
  • The paper lacks direct experimental evidence demonstrating that Noise-Decoupled Chunk Prediction improves inter-chunk continuity. For example, the authors do not report metrics such as inter-chunk action gaps or jerk to directly quantify continuity improvements. Instead, they mainly report success rates, which only provide indirect evidence.
  • Two important benefits of asynchronous inference are (1) hiding inference latency and (2) improving policy responsiveness. However, the evaluated tasks are all static tasks, which do not adequately demonstrate these two benefits.
  • Noise-Decoupled Chunk Prediction requires generating two action chunks, which introduces additional inference overhead and potentially increases inference latency.
  • The experiments investigating the relationship between inference latency and asynchronous execution performance appear to be based on an unreasonable asynchronous inference schedule: once a new inference result becomes available, the system immediately performs trajectory fusion and chunk switching, followed by another inference call. Under this assumption, reducing inference latency naturally increases the chunk-switching frequency, which can lead to less smooth trajectories. However, such a schedule is not necessarily reasonable. A more appropriate asynchronous inference schedule should determine when to initiate the next inference based on the execution progress of the current chunk, or use a dynamic mechanism to decide when a new inference is needed, rather than simply increasing the inference frequency as latency decreases. Otherwise, this may introduce unnecessary inference overhead. Moreover, under a more reasonable asynchronous inference schedule, reducing inference latency would decrease the number of actions executed during inference, thereby alleviating the observation-staleness problem in asynchronous execution. This could instead reduce inter-chunk discontinuities. Therefore, the experimental conclusions drawn from such an unreasonable asynchronous inference schedule raise concerns about their soundness.
Questions for authors
  1. The paper states that Noise-Decoupled Chunk Prediction "generates two action chunks in a single forward pass." Does this introduce additional inference overhead? How much does it affect inference latency?
  2. Could the authors evaluate the proposed method on tasks that require rapid responses to environmental changes?
  3. Could the authors report the improvement in end-to-end task completion time achieved by the proposed method?
  4. Does the asynchronous inference schedule used in this work immediately switch to the newly predicted chunk once an inference call finishes, and then immediately launch the next inference call?
  5. Could the authors provide more details about the real-world experimental setup, including the robot arm model, gripper model, camera configuration, and GPU used for inference?

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

Related papers

Discussion (0)

Sign in to comment.