acceptodds
Under review as a conference paper at ICLR 2027

Reactive Real-Time Flow Policies via Asynchronous Distribution Alignment

Abstract

Generalist robot policies such as vision-language-action models (VLAs) have achieved remarkable generalization, but their inference delays can conflict with the demands of real-time control. Asynchronous execution avoids pauses between action chunks by predicting the next sequence of actions while the robot carries out the previous one. In this paper, we study whether asynchronous execution produces the same action distribution as the original VLA. We find that, for non-Markovian demonstrations, asynchronous execution can produce a fundamentally different action distribution, which can limit the policy’s reactivity. In our method, we seek to restore this reactivity by aligning the asynchronously produced action distribution with that of the original VLA through two complementary mechanisms. First, Recursive Flow-Field Distillation trains the asynchronous policy using the VLA’s action-generation flow. We characterize the learned distribution theoretically and show experimentally that our asynchronous policy can generate nearly the full range of actions the original VLA would produce, while existing asynchronous methods recover only a fraction of that range. Second, Propose–Resolve prepares multiple action sequences asynchronously and uses the latest observation to select among them based on a lightweight approximation of their likelihood under the VLA’s action distribution. Our resulting method matches the original VLA’s success on LIBERO and retains about 80% of its success on RoboMimic, about 30 percentage points more than existing asynchronous methods.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Robot Learning, Real-time Control, Vision-Language-Action Models

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_mGhpOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies why asynchronous execution can produce a different action distribution from the original VLA. It proposes Recursive Flow-Field Distillation (RFD), which trains the asynchronous policy using the frozen VLA’s flow field conditioned on the handoff observation, and Propose–Resolve (PR), which uses the latest observation to select among prepared candidates. Experiments with SmolVLA on LIBERO and RoboMimic examine both action support and closed-loop success, showing better performance over existing asynchronous methods.

Critical strengths

S1. I like the distinction between predicting the demonstrated continuation and reproducing the VLA distribution at handoff. Proposition 1 explains why these targets can differ even with fully observed deterministic dynamics. This gives a clear reason to reconsider what the asynchronous policy should learn. Also, separating expensive proposal generation from lightweight selection is a practical idea. It allows the policy to use the latest observation without generating a new action sequence at handoff. The timing results in Appendix E.4 also provide useful evidence that this separation is computationally feasible with SmolVLA.

S2. The experiments examine both action support and closed-loop success, rather than reporting success rates alone. Applying PR to different proposers helps separate the contributions of proposal generation and selection. The different results on LIBERO and RoboMimic are also informative, showing that broader proposal coverage does not benefit every task equally.

Critical weaknesses

W1. I am still not quite sure what “distribution alignment” means. Appendix C already notes that Gaussian ranking may not preserve the teacher distribution or its mode probabilities. Even with an accurate likelihood estimate, picking the highest-likelihood candidate is different from sampling from that distribution. As the support analysis only looks at proposals before PR, I would like to know whether the final action distribution after selection is also closer to the VLA distribution.

W2. My main concern is that higher success rates may do not directly show that the policy responds better to new information. The resolver uses the latest observation, but I did not find an ablation comparing it with an earlier observation when ranking the same candidates. This would help distinguish the benefit of using new information from the benefit of selecting among multiple candidates.

Questions for authors

Q1. Have you evaluated the final action distribution after PR, and what does “distribution alignment” mean at this stage?

Q2. It is not so clear what RFD adds over TSS in closed-loop execution. Appendix B shows that the two objectives have the same population minimizers under the assumptions, and Appendix D.2 shows better support coverage with RFD. But I didn’t find a comparison of TSS + PR and RFD + PR with the same training budget and number of candidates. Can the authors include TSS in the experiments in Figure 5, using the same training budget and number of candidates? This would help show if the better coverage also leads to better task performance.

Q3. Does using the latest observation work better than using an earlier one when ranking the same candidates?

Please see the Critical Weaknesses section for more details on Q1 and Q3.

3 comments

We thank the reviewer for their recognition of the theoretical motivation and the value of separating proposal generation from selection. We have added the suggested experiments and address their questions below.

1. Latest observation (Q3)

We have added the suggested fresh-versus-stale observation ablation in Fig. 6(a), keeping the proposer, resolver, and candidate count unchanged. On RoboMimic at d=10, ranking with the stale observation achieves 42.3% success, barely above the 41.0% obtained without selection. Ranking with the latest observation instead achieves 58.2%, an improvement of 15.9 percentage points over stale ranking. This clearly shows that the gain comes from incorporating new information, rather than merely selecting among multiple candidates. Appendix E.2, Fig. 13(b), shows how this benefit develops across delays.

2. RFD versus TSS (Q2)

We have added this comparison in Fig. 6(b), applying the same Gaussian resolver to both methods with K=32 candidates. At d=10 on RoboMimic, RFD+PR achieves 58.2% success compared with 51.3% for TSS+PR, a difference of 6.9 percentage points. With a single candidate, both perform similarly: 41.0% for RFD and 41.5% for TSS. The advantage of RFD becomes apparent when selecting among multiple candidates, showing that its improved action coverage can translate into better closed-loop performance. Appendix E.2, Fig. 13(a), extends this comparison across delays.

3. Distribution alignment after selection (Q1)

What does 'distribution alignment' mean? We think this is an interesting and nuanced question, because aligning individual actions with the VLA distribution and reproducing that entire distribution are two different objectives. Ideally, at handoff, we would like to execute an action sampled from the VLA's distribution πVLA(⋅∣o+)\pi_{\mathrm{VLA}}(\cdot \mid o^+). However, obtaining such a sample would require exactly the expensive generative pass that asynchronous execution seeks to avoid. Instead, suppose we already have several candidate actions prepared in advance. The question now becomes: which of these candidates could plausibly have been produced by the VLA at o+o^+?

There are different ways to approach this selection problem. Likelihood-guided (MAP) selection chooses the candidate with the highest likelihood under the fresh VLA distribution. With a perfect likelihood model, sufficient proposal support, and infinitely many candidates, this approaches the MAP action (and our Gaussian resolver approximates this using a diagonal-Gaussian likelihood model and a finite candidate set). We argue that this is still a form of 'alignment' with the VLA distribution: it aims to select an action that the VLA itself would plausibly produce, even though repeated selections will not necessarily reproduce the VLA's sampling distribution.

Alternatively, distribution matching (density-ratio resampling) aims to recover the VLA's actual sampling distribution, including the relative probabilities of different modes. With exact density ratios, sufficient proposal support, and infinitely many candidates, the selected-action distribution converges to πVLA(⋅∣o+)\pi_{\mathrm{VLA}}(\cdot \mid o^+). We had already investigated both approaches but omitted these analyses from the original manuscript for brevity. We have now included them in the revised manuscript, as detailed below.

Distribution after MAP selection. We have added an evaluation of the action distribution after Gaussian selection on 1,000 demonstration handoffs per task (Appendix D.3, Fig. 10). Gaussian ranking increases precision from 0.41 to 0.57 on Square, 0.60 to 0.88 on Transport, and 0.61 to 0.92 on ToolHang, while retaining approximately 73–80% of the original recall. This illustrates how likelihood-guided selection favors actions more consistent with the VLA without necessarily reproducing its full distribution.

Density-ratio resampling. We have added the theory, implementation and results of our density-ratio resolver to Appendix C.2. On held-out demonstration states, it improves precision while retaining 98.2–99.7% of unranked recall (Appendix D.3, Fig. 10). In matched rollout evaluations, however, it achieves 45.5% average success compared with 56.2% for Gaussian ranking (Fig. 7). We therefore retained Gaussian ranking in our method, as it yielded better task performance.

One possible explanation for DRE's lower task success is that estimating density ratios for complex, high-dimensional action distributions is challenging, leaving considerable room for improved estimators.

More fundamentally, this distinction raises an interesting question: Is reproducing the VLA's full action distribution actually desirable for task success, or could consistently selecting its most likely action perform even better? Understanding which objective leads to better closed-loop control, and developing better approximations for both, would be an interesting direction for future work.

Further additions to the revised manuscript

We have also made several additions in response to the other reviews that may be of interest:

  • Additional baselines (Figs. 4–6): We added comparisons with FutureRTC, VLASH, and πR² to contextualize our approach against recent developments in asynchronous VLA execution.
  • Candidate-count ablation (Fig. 6c): Substantial improvements already emerge with K=4 candidates, with diminishing returns at larger candidate counts.
  • Extended PR comparisons (Appendix E.2, Fig. 12): We compare compatible proposers across delays using the same resolver and K=16 candidates.
  • Clearer distribution visualization (Fig. 3): Separate PCA panels compare each method with the original VLA to make differences in coverage easier to distinguish.

Thanks a lot for the detailed response! The revised manuscript looks good, and I will keep my positive score.

Reviewer_VqiyOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a method to improve generalist robot policies from VLAs in asynchronous execution. Given that human demonstrations are non-markovian and that in those situations asynchronous policies output a reduced range of possible responses, it investigates how to align with the original VLA policy. The paper proposes two steps to partially recover and use the full action distribution: First to learn the action distribution of VLAs given possible immediate future observations and then propose candidate actions during execution of the current action chunk. Second, to choose the best of these candidates at the handoff to predict the next chunk. Ranking is realized with a diagonal-Gaussian approximation given the observation at handoff time.

Critical strengths

The general idea of developing plausible actions during execution and then ranking the result for fast execution is very interesting and similar to speculative decoding.

The papers proposal of mismatched action distribution is illustrated with a strong formal argument, showing how the mismatch between the VLA policy action distribution and an asynchronous policy action distribution is unavoidable for non-markovian demonstrations. In a very principled manner the issue is approached with the proposed speculative action generation.

The performance compared to the asynchronous baselines is very impressive, taking recall and precision together and given some issues with one-sided performance. Resolving candidates in 12 ms seems like a fair trade-off.

Distilling the velocity field for the RFD policy is an interesting alternative to distilling samples. Learning the VLA's field allows to cover all potential actions.

The paper is transparent about many limits e.g. of the Gaussian scoring.

Critical weaknesses

The rationale highlights issues with large slower VLAs but SmolVLA is not a large model and, as mentioned in the limitations, the resolver still needs a forward pass. It would have been interesting to see a comparison on another larger flow-based VLAs.

Precision seems to suffer a lot even on demonstration states compared to all other approaches as seen in figure 6 and 4. That may be an issue as the goal is to improve reactivity but in unexpected situations where reactivity is important the performance is worst.

The training configuration and Table 2 shows that each configuration uses a single training seed. A few more seeds could have been helpful to report variance of the results. OpenVLA used three seeds which seems like a reasonable minimum.

Minor point: The trajectories in Figure 3 seem very different. Seeing those as qualitative examples in videos for the supplementary material could serve well to see the difference between approaches.

Questions for authors

What is the interpretation of the low performance in precision? Precision and recall are often inversely correlated. Do we see this effect here? Is the approach very sensitive to larger d values in general and limits its action distribution on rollout states more heavily as d increases than other approaches?

How will this approach transfer to larger or slower models when d may increase even more?

Did you investigate alternatives to the Gaussian selection which may be slower and if yes, did you investigate which trade-off is possible in terms of selection time vs. improved performance? This is mentioned as a limitation but seems like a noteworthy information for the appendix when people want to build on this.

For the square case where the recall drops almost to 0, what is the performance/contribution of the proposer and the resolver?

3 comments

We thank the reviewer for their encouraging assessment, particularly for recognizing the strength of the formal analysis and the value of preparing alternative actions before resolving them at handoff. Below, we address the reviewer's questions and summarize our changes to the manuscript.

Additional experiments and appendix material

1. Alternatives to Gaussian selection.

We had also investigated density-ratio resampling as an alternative that targets recovery of the fresh VLA distribution rather than selecting the highest-scoring candidate. We agree that these findings may be useful to others building on our work, so we have now included the formulation, implementation, and results in Appendix C.2. On held-out demonstration states, DRE improves precision while retaining 98.2–99.7% of unranked recall (Appendix D.3, Fig. 10). In separate matched rollout evaluations, it increases average success from 40.2% to 45.5%, compared with 56.2% for Gaussian ranking (Fig. 7). Thus, this alternative preserves more action diversity but does not outperform Gaussian ranking in task success.

Regarding selection time, DRE also uses lightweight neural-network scoring, although we did not separately benchmark its latency. We did not investigate whether more expensive resolvers would improve performance, since selection must complete within a single control step. Otherwise, the additional computation would quickly undermine the real-time advantage of resolving prepared candidates rather than generating a new VLA action.

Clarifications to the reviewer

2. Square: proposer and resolver contributions.

In the Square evaluation at d=10, RFD alone achieves 48% success, increasing to 55% with Gaussian PR over K=32 candidates (Appendix C.2, Fig. 7). Thus, both the proposer and the complete method retain useful task performance, even though proposal recall drops to nearly zero on VLA rollout states (Appendix D.1, Fig. 8). One possible explanation is that Square has a particularly narrow VLA action distribution, where even small deviations in the learned proposals can lead to near-zero measured recall, without necessarily making those actions ineffective for the task.

3. Precision, recall, and increasing delay.

Why is precision lower? RFD generates candidates before the handoff observation is available. As characterized by Eq. 6, it must therefore average over the VLA distributions associated with possible future observations. This preserves a broader range of possible actions (high recall), but also includes actions that may be inappropriate for the observation actually reached (lower precision).

Are precision and recall inversely correlated? To some extent, yes, but this trade-off is not universal. More importantly, our full method does not directly execute the low-precision RFD proposal distribution. Instead, PR uses the fresh observation to select appropriate candidates, substantially increasing precision while preserving much of the original recall. Our new post-selection analysis (Appendix D.3, Fig. 10) shows precision increasing to 0.88 on Transport and 0.92 on ToolHang, while retaining approximately 73–74% of the original recall. Thus, the complete method achieves a much better balance between coverage and precision than the raw proposals alone.

Does performance degrade more strongly with increasing delay? Not generally. RFD maintains high recall across delays on demonstration states. On rollout states, recall remains strong on ToolHang and degrades moderately on Transport, although Square presents a clear outlier (Appendix D.1, Fig. 8). However, precision and recall measure overlap with the VLA's action distribution, not whether the proposed actions actually succeed at the task. As the Square example illustrates, even near-zero measured recall can coexist with substantial task success. Our closed-loop results (Fig. 5) are therefore more informative about robustness to delay: RFD+PR maintains strong performance, especially as dd increases.

4. Transfer to larger VLAs.

Larger VLAs generally require more time to generate actions, which is exactly the problem asynchronous execution is designed to address. As inference time increases, we can increase the delay dd to prepare actions further in advance. At 20 Hz, d=10d=10 and d=20d=20 correspond to inference times of 500 ms and 1 second, respectively. Even at d=20d=20, RFD+PR achieves 60.7% success on LIBERO, compared with 62.5% for the synchronous VLA (Fig. 5), demonstrating robustness to substantial inference delays. Moreover, RFD's flow-based formulation is not tied to SmolVLA's model size, and PR selects among prepared candidates without requiring a new generative pass. We therefore see promising potential for applying our approach to larger VLAs. The main practical requirement is that feature extraction and resolution can still be completed within a single control step, which we leave for future investigation with larger models.

Further additions to the revised manuscript

We have also added several experiments in response to the other reviews that may be of interest:

  • Additional baselines (Fig. 5): We added comparisons with FutureRTC, VLASH, and πR² to contextualize our approach against recent developments in asynchronous VLA execution.
  • Candidate-count ablation (Fig. 6c): We show that substantial improvements already emerge with K=4 candidates, with diminishing returns at larger K.
  • RFD versus teacher-sample supervision (Fig. 6b): While both achieve similar single-proposal performance, RFD+PR achieves higher success (58.2% versus 51.3% on RoboMimic), showing that improved proposal coverage can translate into better execution.
  • Fresh versus stale observation (Fig. 6a): Ranking with the latest observation achieves 58.2% success versus 42.3% with the stale observation on RoboMimic, demonstrating that the benefit comes from incorporating new information rather than merely selecting among candidates.
  • Qualitative visualization (Fig. 3): Rather than adding videos of the short candidate trajectories, we reworked the figure into separate PCA panels comparing each method with the VLA. This makes the intended comparison of action distributions clearer.

The authors have addressed most of my concerns, thank you very much for adding additional baselines. I will increase my score for now and have a positive impression going into the final reviewer-internal discussion phase.

Reviewer_6GJPOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper address the distribution mismatch between the prefix-conditioned policy with stale observation and the true distribution of the pretrained VLA, under asynchronous inference. To tackle this problem, the paper proposed 2 components: Recursive Flow-Field Distillation (RFD) and Propose-Resolve (PR). RFD enforces an asynchronous student, observing (o-, P), to predict d steps ahead to match the distribution of the pretrained VLA teacher, taking latest observation o+ as input. However, as RFD reconstructs the teacher distribution, it marginalizes over all reachable observations from the proposed distribution, forming a broader distribution compared to the original one. To tighten this broad distribution, PR serves as a filter to select the best action from K candidate using a small solver, which maximize action likelihood under the true VLA distribution. The paper demonstrated the effectiveness of this approach both theoretically and empherically with the RoboMimic and LIBERO suites, outperforming 3 baselines in terms of recall-precision trade-off and retaining 80% to 100% performance of the pretrained model depending on tasks. The ablation study also highlight how the contribution of each component changes depending on tasks.

Critical strengths
  • The method is technically sound and well-motivated for the distribution mismatch problem, preventing performance loss while keeping the pretrained model frozen. The paper provides a clear explanation of the Recursive Flow-Field Distillation (RFD) and Propose-Resolve (PR) components. RFD aligns the (o-, P) policy with the o+ policy by distilling the teacher's velocity field. PR then narrows the resulting broad distribution with the real o+. The two components work together to address the distribution mismatch effectively.
  • The paper provides insightful analysis on component contribution. The K=1 vs K=32 ablation (Fig. 5, bottom) and the discussion of when distribution matching matters (RoboMimic) versus not (LIBERO) are candid. The paper is well-structured and easy to follow, with the proposed methods clearly explained and justified. The theoretical analysis provides necessary background for the proposed approach.
Critical weaknesses
  • In the ablation for PR contribution in figure 5, PR significantly boosts the performance of RFD and seems to a more critical component than RFD, but it can't handle multimodal policy due to the simple architecture by construction. For more complex tasks, it is questionable whether the proposed approach can carry over, which limits its applicability to complex tasks like long-horizon of bimanual. Moreover, for PR design, the practical cost and scalability of this approach for real-world deployment is not discussed. Preparing K=32 candidates costs about 91 ms on an RTX 4090, which fits only if d is large enough. With smaller d, the cost of preparing K candidates may outweigh the benefits of PR.
  • Comparison to closely related concurrent work is missing. The related work notes that other methods predict execution-time state or visual features (VLASH[1], FutureRTC[2]) or incorporate newly measured state during generation (piR^2[3]), but none of them is compared empirically. Given that distribution alignment does not directly transfer to success rate, it would be helpful to see on the high level, which direction is more effective for asynchronous inference.
  • Again, experiments are limited to the simulation. It would be beneficial to see how the proposed method performs in real-world scenarios, especially in tasks that involve more complex dynamics and interactions where real latency jitter, camera delay, etc exacerbate the inference delay.

Minor issues

  • Fig. 3: add PAINT and consider using a more distinguishable palette, the current color palette make it hard to distinguish baselines.
  • Figs. 4 and 5: bold the best result in each column (including precision) consistently.

[1]https://arxiv.org/abs/2607.24008 [2]https://arxiv.org/abs/2512.01031 [3]https://arxiv.org/abs/2607.26055

Questions for authors
  • Can you show Fig.5 with PR applied to all baselines at the same K=32 across different d? As the current d and K is selected in favor of the proposed method while PR holds a lot of potential for improving performance, it would be interesting to see if other baselines can also benefit from a suitable pair of d and K.
  • How does RFD+PR compare to FutureRTC [1], VLASH[2] or piR^2[3] under the same backbone and delays? This may help to understand the trade-offs between different approaches for asynchronous inference, especially when distribution alignment does not directly convert to higher performance.
  • How does performance depend on K? Is K=32 the optimal choice, or is there a trade-off between K and performance?
  • Is there any real-robot result? How's the method affected by real-world factors like latency jitter, camera delay, etc?
  • What is the setting for synchronous SmolVLA inference in this paper? Why the success rate of the frozen model lower than reported in the original paper ([4] Table 2)? Is it due to the difference in the evaluation setting or implementation details?
  • Why there are two different ranges for d (5-10-15-20 vs 5-6-7-...10) in Fig 5 for the two benchmarks? Can you provide the full result for this ablation? (2 task suites x 2 ranges)

[1]https://arxiv.org/abs/2607.24008 [2]https://arxiv.org/abs/2512.01031 [3]https://arxiv.org/abs/2607.26055 [4]https://arxiv.org/pdf/2506.01844

3 comments

We thank the reviewer for the encouraging assessment of our work, particularly for recognizing the clarity of its presentation and the insights provided by our analysis. We appreciate the suggestions for extending this analysis, many of which we agree would make the paper stronger. We uploaded a revised manuscript with several new experiments added. We discuss our additions, along with the remaining questions, below.

New experimental results and manuscript revisions

1. Added comparisons with FutureRTC, VLASH, and πR².

FutureRTC and πR² appeared on arXiv only in late July 2026, leaving a short window before our submission for controlled empirical comparisons. We therefore initially discussed them as concurrent work, consistent with ICLR's guidelines. Nevertheless, we agree with the reviewer that these comparisons provide valuable context and have now added evaluations of all three methods to Section 5.

RFD continues to provide the broadest proposal coverage in most settings, while FutureRTC is competitive and also achieves high precision and strong performance at K=1K=1. Figure 5 shows that RFD+PR achieves the highest success across the tested LIBERO delays and at larger delays on RoboMimic. Interestingly, Figure 6 reveals that these approaches are complementary: RFD+PR performs best on RoboMimic, whereas FutureRTC+PR and VLASH+PR lead on LIBERO. This demonstrates that Propose–Resolve is not limited to RFD, but can also substantially improve other asynchronous policies. More broadly, these findings reinforce the question of which continuation distribution is most useful for asynchronous execution, and how different proposal and selection mechanisms can best be combined.

2. Added an ablation over the number of candidates KK.

We added an ablation in Fig. 6(c), which shows that while K=32K=32 achieves the highest success, substantial gains already emerge at K=4K=4, with diminishing returns thereafter. Importantly, this also addresses the concern about computational cost: preparing four candidates takes only 33 ms (Table 3, Appendix F.4), making the approach practical even for smaller inference delays. Appendix E.2 further shows the candidate-budget comparison across delays.

3. Extended the PR ablation across delays.

Furthermore, to show how the benefits of PR evolve across different inference delays, we extended the RoboMimic ablation to all compatible proposers across d=1,…,10d=1,\ldots,10 (Appendix E.2, Fig. 12). We apply the same Gaussian resolver with K=16K=16 candidates to every proposer, including RFD, ensuring a direct comparison under matched conditions. Note that we use K=16K=16 instead of K=32K=32 to reduce evaluation costs, as generating multiple candidates is particularly expensive for methods such as RTC and PAINT due to their generative procedures.

Clarifications to the reviewer

4. Different delay ranges.

Real-time chunking requires the inference delay dd to satisfy d≤sd \leq s, where ss is the executed chunk length. On RoboMimic, we use s=10s=10, as SmolVLA's performance degrades for longer execution horizons. On LIBERO, we use s=25s=25 to investigate robustness to substantially larger inference delays. Our original RoboMimic evaluation focused on d=5,…,10d=5,\ldots,10, where differences between methods become more pronounced. To avoid confusion and provide a more complete comparison, we have now expanded the main results (Fig. 5) to cover the full range d=1,…,10d=1,\ldots,10.

5. Synchronous SmolVLA performance.

Our evaluation uses a longer executed action horizon (s=25s=25 on LIBERO), whereas the original SmolVLA headline results use an execution horizon of 1. The original paper's own ablations (Table 13) also show that increasing ss leads to substantial performance degradation, consistent with the performance we observe under our longer-horizon evaluation setting.

6. Multimodal policies.

We would like to clarify an important distinction: PR is not inherently incompatible with multimodal policies. The Gaussian resolver does not generate actions; it only ranks candidates generated by the flow-based proposer, which can represent multiple modes. Thus, the Gaussian approximation does not restrict the proposed actions to a single mode. It can certainly lead to imperfect rankings, as we discuss in Appendix C.1, but this is different from being unable to handle multimodal policies by construction. In fact, RoboMimic Transport is a long-horizon, bimanual manipulation task on which our method achieves strong results. We have also added an investigation of density-ratio resampling as an alternative resolver in Appendix C.2.

7. Real-world evaluation.

Our primary objective is to study the action distributions induced by different asynchronous continuation methods. Simulation allows us to compare these distributions at identical handoff states and evaluate their closed-loop behavior from reproducible initial conditions. This level of control is considerably harder to achieve in the real world, where variations between resets can obscure differences between methods.

Regarding deployment, inference jitter and additional latency during proposal preparation can be accommodated by increasing the prediction horizon dd, allowing more time for candidates to be generated before handoff. At handoff, our resolver requires approximately 12 ms on an RTX 4090, leaving roughly 38 ms within a 50 ms control period for camera acquisition, communication, and other overhead (Appendix F.4). We do not currently have real-robot results, and evaluating the complete pipeline under realistic timing conditions remains valuable future work.

Minor manuscript changes

8. Presentation improvements.

We have reworked Fig. 3 to show each baseline individually against the original VLA in separate PCA panels, making differences in distribution coverage easier to distinguish. We also added PAINT and removed the candidate trajectories, as they conveyed largely the same information while making the visualization harder to interpret. Finally, we made best-result highlighting consistent across Figs. 4–5.

Further additions to the revised manuscript

We have also added several experiments in response to the other reviews that may be of interest:

  • Fresh versus stale observation (Fig. 6a): Ranking with the latest observation achieves 58.2% success versus 42.3% with the stale observation, demonstrating the importance of incorporating new information at handoff.

  • RFD versus teacher-sample supervision (Fig. 6b): RFD+PR achieves higher success than TSS+PR (58.2% versus 51.3%), showing that improved proposal coverage translates into better execution.

  • Post-selection distribution analysis (Appendix D.3, Fig. 10): Gaussian ranking substantially improves precision while preserving much of the original recall.

  • Alternative resolution method (Appendix C.2, Fig. 7): Density-ratio resampling better preserves action diversity, but achieves lower task success than Gaussian ranking (45.5% versus 56.2%).

Dear Authors,

Thank you for your new experiments and clarifications for the paper. They effectively address most of my concerns. I also appreciate how authors quickly adopted new baselines for a more comprehensive evaluation. It would be even better if real-world experiments were provided but in general, I'm happy to raise my score. Great work!

Comments on the paper

1 comment

Hi everyone, please do acknowledge the authors’ rebuttal with a brief reply, even if your assessment is unchanged (if you already did or are about to do so, please ignore this message).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.