- 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
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.
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+). 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+?
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+). 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.