acceptodds
Withdrawn from ICLR 2027

RAVEL:Asynchronous Rolling Inference for Flow-Based Vision–Language–Action Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Flow-based vision–language–action (VLA) models are highly effective for gen- eralist robot manipulation, yet their reliance on computationally expensive VLM encoding and multi-step iterative action generation imposes a significant latency bottleneck. The resulting inference latency makes it difficult for robots to respond quickly, especially in dynamic environments. We address this limitation with RAVEL (Rolling Asynchronous VLA Enabling Low-Latency Control), an asyn- chronous inference framework that addresses the computational bottlenecks of both the VLM backbone and the action expert. To reduce the delay from multi-step action denoising, RAVEL allows near-term actions to be executed after a single denoising step by carrying partially denoised future actions forward in a rolling buffer. To avoid blocking on slow VLM encoding, RAVEL decouples VLM encoding from rolling action generation, allowing the action expert to operate continuously using the latest available VLM context, while a lightweight Fast Observation Pathway (FOP) directly conditions the action expert on current obser- vations. Across simulated and real-world manipulation tasks, RAVEL consistently achieves substantially lower response latency while maintaining the task capabil- ity of the underlying VLA, enabling high-frequency and responsive closed-loop control.

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

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_pGNLOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

RAVEL adapts pretrained flow-based VLAs to asynchronous control. It maintains a rolling buffer with position-dependent noise levels, progressively refines future actions, and releases one buffer slot per action-expert update. VLM encoding runs independently; the expert receives the latest available semantic context and its temporal offset. An optional Fast Observation Pathway (FOP) supplies current visual and proprioceptive features. Training pairs older observations with shifted action horizons and matches the rolling noise profile. Evaluation covers LIBERO, RoboTwin, DOM, and two physical tasks.

Critical strengths
  1. The framework addresses a practical bottleneck and is demonstrated on multiple pretrained backbones. Median first-action latency decreases by 2.59x for X-VLA and 2.92x for pi0.5 while synchronous benchmark performance is largely preserved.
  2. Explicit context-age conditioning is a useful, identifiable addition. With the checkpoint and runtime fixed, using the correct offset achieves 44.89% DOM success versus 16.67% when the offset is reset upon receiving new context.
  3. Dynamic results are promising: RAVEL reaches 44.06% DOM success versus 35.50% for the best listed baseline, with FOP increasing this to 46.94%. The appendices explain implementation choices, training, and timing limitations.
Critical weaknesses
  1. The broad architectural novelty is limited. HiRT already combines a low-frequency VLM, a high-frequency visual policy, asynchronous reuse of semantic features, and training with older visual inputs (HiRT, Secs. 3.2-3.3; CoRL 2024). Thus, decoupling perception from control, bypassing the VLM with current visual features, and training for stale context are established ideas. RAVEL's stronger candidate contribution is their integration with rolling flow inference and explicit temporal-offset conditioning. The related-work discussion does not sufficiently explain what new knowledge this integration establishes.
  2. The experiments do not isolate this candidate contribution against a conventional fast/slow architecture using the same backbone. Sec. 4.4 holds rolling inference fixed and ablates its training components; the DOM FOP comparison only establishes FOP's incremental benefit within RAVEL. Neither tests asynchronous context reuse plus current-observation conditioning without rolling, or rolling with synchronous context updates, under comparable adaptation.
  3. Empirical attribution remains uncertain. DOM variants use different training scopes, batch sizes, and checkpoint-selection procedures (App. A.2); synchronous baselines do not receive equivalent extra adaptation (App. A.1). More training steps alone do not imply more data exposure. In table tennis, RAVEL additionally uses a tuned offset bias of 0.3, corresponding to 100 ms of lookahead for physical tracking lag. Equivalent execution-delay calibration for competing methods is not reported (Secs. 4.3-4.4; App. B.4).
  4. Real-world evidence is narrow: 20 trials per method on one robot, and grasping success equals naive asynchronous inference at 45%. DOM's completion-time comparison uses only 36 episodes successful for all six methods, rather than the full 1,800-episode benchmark. These limitations temper claims of broad practical superiority.
Questions for authors
  1. On one existing dynamic benchmark, can you compare asynchronous context reuse plus FOP with conventional full-chunk denoising against the rolling variant, using comparable data exposure and adaptation? A rolling variant with synchronous context updates would further clarify the two mechanisms' interaction.
Reviewer_HMAhOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose RAVEL, an asynchronous inference framework for flow-based vision-language-action models. The method combines a rolling action buffer with asynchronous VLM encoding to reduce the latency of action generation. Instead of waiting for multiple denoising steps before executing an action, RAVEL progressively refines future actions and releases the next action after a single denoising update. The authors also introduce temporal-offset conditioning and a Fast Observation Pathway to incorporate more recent observations while reusing previously computed VLM features. Experiments on LIBERO, RoboTwin 2.0, and Dynamic Object Manipulation show improved inference responsiveness while largely preserving task performance. Real robot experiments further demonstrate improved table-tennis return success and smoother manipulation trajectories.

Critical strengths

The paper addresses an important practical limitation of current VLA models, particularly the latency introduced by expensive visual encoding and iterative action generation. I appreciate that the authors evaluate responsiveness in dynamic environments rather than focusing only on inference speed. The experiments cover multiple VLA backbones and include both simulation and physical robot evaluations. The ablation studies on temporal-offset conditioning are also useful and show that accounting for observation delays is important for maintaining performance during rolling execution. The reported improvements on dynamic manipulation tasks are encouraging.

Critical weaknesses

My main concern is the limited methodological novelty. The authors propose combining rolling action denoising with asynchronous visual encoding, but both directions have already been extensively explored in the literature. Responsive Noise-Relaying Diffusion Policy and Rolling Diffusion Policy introduce closely related rolling denoising mechanisms, while Streaming Flow Policy studies continuous action generation during flow sampling. More importantly, Reflex already investigates asynchronous VLA inference by decoupling visual encoding from action generation. I find that RAVEL mainly integrates these existing ideas with additional temporal conditioning and a lightweight observation pathway, rather than introducing a fundamentally new inference mechanism. Although the authors discuss these related directions, direct comparisons with some of the closest methods are missing. My second concern is the fairness and interpretation of the experimental comparisons. In Table 2, several strong baselines, including VLASH and FASTER, perform worse than the original DynamicVLA model. The authors acknowledge in Appendix A.2 that their adaptations may suffer from implementation and action representation mismatches. This makes it difficult to determine whether the reported improvements primarily reflect the proposed method or weaknesses in the baseline implementations. The real robot comparison raises similar concerns. RAVEL achieves 90% success on table-tennis return, but this result uses an empirically selected temporal-offset bias of 0.3, while the other methods do not receive an equivalent execution-delay compensation mechanism. Moreover, the evaluation contains only 20 trials per method, and RAVEL receives additional training compared with the original synchronous and naive asynchronous policies. My third concern is the limited evidence supporting the claimed real-time efficiency. The Fast Observation Pathway adds nontrivial computation, and the reported action-expert latency with FOP reaches 41.4 ms on an RTX 4090, which already exceeds the 40 ms control interval of the 25 Hz simulation environment. The paper does not sufficiently analyze scheduling failures, missed control deadlines, or the full computational overhead of the asynchronous pipeline. Overall, I find the engineering results interesting, but the novelty and experimental evidence are not strong enough to support acceptance.

Questions for authors

Could the authors clarify the fundamental technical differences between RAVEL and existing rolling denoising or streaming inference methods, particularly Responsive Noise-Relaying Diffusion Policy, Rolling Diffusion Policy, Streaming Flow Policy, and Reflex? I would appreciate a more direct comparison under matched models and computational budgets. Second, the reproduced VLASH and FASTER results on DOM are considerably weaker than the baseline DynamicVLA model. Can the authors explain whether these results reflect limitations of the original methods or difficulties in adapting their implementations? Third, for the real robot experiments, could the authors evaluate the baselines with comparable execution-delay compensation or provide an ablation that separates the contribution of the temporal-offset bias from the rolling inference framework? Finally, can the authors report actual control frequencies, missed deadlines, and end-to-end latency distributions with FOP enabled? These results would help determine whether RAVEL consistently supports real-time execution rather than only reducing average inference latency.

Reviewer_fX3SOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

RAVEL lets a VLA keep producing actions while its VLM processes the next image. At each control step, the action expert refines a buffer of upcoming actions, sends the first action to the robot, and keeps the rest for the next update. It uses the age of the latest VLM context to predict actions for the intended execution time. FOP can also supply the expert with the current image and robot state. RAVEL largely maintains task success on LIBERO and RoboTwin 2.0 while reducing first-action latency. It improves success on DOM and real-world table-tennis returns. On tennis-ball grasping, it matches the best baseline’s success rate with smoother motion.

Critical strengths
  1. The paper addresses a practical timing issue. Even when a VLM context has just become available, the image it represents may already be several control steps old. RAVEL accounts for this delay during both training and inference. The DOM ablation shows that the policy performs better when it measures the delay from image capture, including the time spent on VLM encoding.

  2. Table tennis is a good test of the method because a well-aimed swing can still miss if it comes too late. RAVEL returned substantially more balls than the other methods tested on the same robot. This gives the latency results a clear practical meaning: the complete system can respond in time to hit a moving target.

Critical weaknesses
  1. It is unclear how much RAVEL gains from repeatedly revising upcoming actions. A simpler policy could also use the age of the image to predict an action chunk for the correct time. The ablations in Sec. 4.4 keep rolling inference in place, so they show why timing information helps RAVEL but leave open how well that simpler policy would perform. This comparison matters because revising pending actions at every control step is central to the method.

  2. The table-tennis result is sensitive to the delay between sending a command and the arm reaching its target. Fig. 5(c) shows that changing how far ahead RAVEL predicts has a large effect on return success, even with the same policy weights. The paper reports the inference delays assumed by VLASH and FASTER, but I could not tell how either method accounted for the arm’s movement delay. This makes the size of RAVEL’s lead in Table 3 difficult to interpret.

Questions for authors
  1. On DOM, how does RAVEL compare with a policy that uses the same asynchronous VLM, image-age information, and training budget, but generates conventional action chunks? Its success rate and response time would help clarify what rolling refinement adds.

  2. How was RAVEL’s lookahead selected for Table 3? Were the reported trials separate from those used for the sweep in Fig. 5(c)? Could VLASH and FASTER compensate for the same physical delay, and if so, how were they configured?

I lean weak reject. The motivation is clear and the real-world results are promising. I would consider raising my score if the authors show how much rolling refinement helps over a simpler policy that also accounts for delay.

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

Related papers

Discussion (0)

Sign in to comment.