acceptodds
Withdrawn from ICLR 2027

Beyond Noise: Regulating How Much of the Prior Survives in Flow-Matching Robot Policies

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Most flow-matching robot policies generate actions from Gaussian noise, ignoring the temporal continuity available in robot interaction. Recent methods warm-start the flow with informative sources such as previous actions or visual forecasts, implicitly assuming that a better source produces a better action. We show that this assumption is incomplete: informative sources help, but much of their information is attenuated by the flow before it reaches the executed action, so what must be controlled is how much of the prior survives, not only which prior enters. We therefore propose Beyond Noise. It enters the flow at a state that couples the prior’s weight, the injected noise and the remaining path through a single entry time q. Beyond Noise selects a default entry on validation data and applies the regulator’s per-call refinement where validated, dynamically adjusting q according to the current state; the regulator learns from counterfactual closed-loop returns. A patch-token memory lets the policy use where objects are once the prior is kept. On four LIBERO suites, under a protocol that makes every choice on held-out initial states, Beyond Noise reaches 81.0% average success with a single function evaluation, above matched Gaussian, visual-prior and external baselines, with the largest gains on long-horizon tasks. Exploiting a better prior requires deciding how much of it survives.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Flow Matching, Generative Robot Policies, Robot Manipulation, Closed-Loop Control, Temporal Priors, State-Conditioned Source Regulation

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_AafnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how action priors can improve flow matching policies for robotic manipulation. The authors propose a method that initializes the flow using a combination of a separately predicted action and Gaussian noise, while jointly adjusting their relative weights and the remaining integration interval through a single parameter. The flow is trained to accommodate different initialization settings, and a default parameter is selected through validation, with a learned regulator making state-dependent adjustments where validated. Additional spatial visual features are incorporated to improve action refinement. Experiments on LIBERO benchmarks demonstrate improved overall performance over existing methods, supported by ablation studies and robustness evaluations.

Critical strengths

The paper provides a systematic analysis of how different flow initializations affect action generation and closed-loop performance, offering useful empirical insights into the role of action priors. The proposed method is intuitive and relatively simple in its formulation, using a single parameter to jointly control the contribution of the predicted action, noise level, and integration interval. The experimental results demonstrate overall improvements, particularly on challenging long-horizon manipulation tasks.

Critical weaknesses

Existing flow policies are already conditioned on the current observation through their learned vector fields. While this paper additionally incorporates an action prior from a separately trained predictor into the flow initialization, the distinction between these two ways of incorporating information, and their respective advantages and limitations, are not sufficiently discussed. The proposed method also introduces additional computational overhead despite using a single flow step, which other flow policies can also achieve.

The final method combines several components, including specialized training, patch-token memory, and adaptive regulation. Although the paper provides several ablation studies, the interactions between these components are not fully isolated, and the benefits of coupling noise scaling with flow entry time remain unclear. Furthermore, the regulator is trained using fixed-entry continuations, whereas the deployed policy may adjust the entry time at each planning step.

I found two potentially inaccurate citations in the Related Work section:

  1. Sheng et al. (2026), MP1 (p. 9, lines 461–462): it is categorized as a distillation method, but it directly trains a policy using MeanFlow rather than distilling a pretrained teacher.
  2. Liu et al. (2026), Learning native continuation for action chunking flow policies (p. 9, lines 473–475): it is grouped with methods that enter the flow at an intermediate timestep, but it uses continuation guidance throughout the denoising process rather than starting partway through the flow.

I think these two references should be reclassified or their descriptions revised.

Questions for authors
  • Could the authors clarify whether the benefits of incorporating an action prior through flow initialization justify the additional computational overhead compared to existing observation-conditioned one-step flow policies?
  • Given that noise scaling appears to have the largest effect on prior transmission, could a simpler approach that adaptively adjusts only the noise scale, without coupling it to the anchor weight and entry time, achieve comparable performance?
  • How reliable are the regulator's estimated values when the entry time is adjusted at each planning step, rather than held fixed throughout the continuation?
Reviewer_aJEeOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how informative starting distributions affect flow-based robot policies. The authors find that warm starts improve performance, but much of the difference between starting predictions is minor before reaching the final action. They propose Beyond Noise, which jointly controls the base action chunk and noise level through a scalar parameter. This parameter is selected using validation rollouts and optionally adjusted during execution by a regulator trained on counterfactual outcomes. The method also incorporates image patch features and reports improvements on LIBERO, particularly on long-horizon tasks.

Critical strengths
  • The paper investigates the question of whether a more informative starting prediction produces a better flow-based policy and provides experimental studies beyond overall success rates.

  • The authors make an effort to control evaluation noise and separate model selection from testing. The paper also reports cases where adaptation does not help.

Critical weaknesses
  • It is unclear whether the proposed coupling is necessary. The ablations suggest that reducing noise matters most, while changing the flow entry time has little effect on success.

  • The regulator is not described in enough detail. It is unclear how the value model is trained, how rollout starting points are chosen, and how much simulator interaction is required.

  • Some experimental settings are not described clearly. For example, the proposed method adds image patch features and extra training. Since these changes help substantially, the experiments should separate their benefits from those of the proposed entry-time mechanism.

Questions for authors
  • In Eq. 4, why does rescaling the loss leave the optimal solution unchanged when qq varies across training examples? Does the model receive qq as a separate input?

  • How is the regulator trained, and how many simulator rollouts does it require?

  • Could the authors separately ablate the gains from patch features and choosing a fixed entry time?

  • Would simply reducing the noise in a standard warm-start policy achieve similar performance?

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

The paper studies how informative initial sources affect flow-matching robot policies. A matched-source study finds that warm starts improve closed-loop performance, while differences between informative sources and perturbations to them are substantially attenuated in the generated actions. The proposed Beyond Noise method constructs Xq=qBt+(1−q)ϵX_q=qB_t+(1-q)\epsilon, starts integration at flow time qq, and trains the policy on these entry states. It combines validation-selected default entries with a state-level regulator trained from counterfactual closed-loop continuations, enabling per-call refinement where validation supports it. Patch-token conditioning provides additional spatial information. Experiments are entirely simulated, covering four LIBERO suites, execution perturbations, and LIBERO-plus. The complete system reports 81.0% average success, compared with 73.3% for the fixed visual-prior baseline.

Critical strengths
  1. The matched-source investigation provides useful empirical observations. The comparisons of Gaussian, temporal, and visual sources, together with the consistency and source-perturbation analyses, go beyond reporting task success alone. In particular, the distinction between offline prediction error and agreement between consecutive plans is useful for understanding warm-started policies. This is the strongest informative aspect of the submission.

  2. The complete system produces substantial gains on the harder simulated tasks. The improvement over the fixed visual-prior baseline is particularly large on LIBERO-10, from 44.0% to 64.3%, with an additional improvement on Goal. These suites do not use the additional moved-object demonstrations introduced for Object and Spatial, so the data-augmentation concern below does not explain away all of the reported gains. The validation/test separation and additional fresh-state evaluation are also worthwhile controls.

Critical weaknesses
  1. The proposed coupling mechanism lacks convincing evidence of its distinctive contribution.

The central claim is that coupling prior weight, noise scale, and flow entry time improves policy performance. However, Section 5.3 reports that reverting to s0=0s_0=0 leaves success unchanged, while noise scale is identified as the dominant factor. Moreover, higher prior transmission does not consistently improve success. These findings suggest that reduced noise, rather than the proposed three-way coupling, may explain much of the improvement.

The benefits of state-dependent regulation are also less general than suggested. On clean LIBERO, per-call refinement is enabled only for LIBERO-10, while the other suites use fixed defaults. Furthermore, Appendix F reports that the preregistered primary refinement rule achieves a statistically insignificant improvement (p=0.29p=0.29). The evidence therefore does not sufficiently establish the distinctive value of the proposed mechanism.

  1. Experimental limitations and insufficiently controlled comparisons undermine the reported improvements.

Appendix A.3 reveals that Beyond Noise uses 1,742 and 1,563 additional expert demonstrations on LIBERO-Object and Spatial, respectively, without clearly establishing whether the baselines receive equivalent data. These demonstrations substantially affect performance: Appendix G shows that clean Spatial success increases from 66.9% to 79.4%, while object-shift success improves from 29.6% to 52.0% on Object. Combined with the additional patch-token memory and specialized training curriculum, this makes it difficult to isolate the contribution of the proposed regulation mechanism.

More importantly, the paper provides no real-world robotic experiments. All results are obtained in simulation, despite the emphasis on robust closed-loop control. Whether the method remains effective under real-world perception errors, actuation uncertainty, and control latency is entirely unverified. This is a substantial limitation for a submission in the Applications to Robotics, Autonomy, Planning area.

Questions for authors
  1. Relationship to MARS Policy. A closely related preprint, MARS Policy: Multimodality Only When It Matters (arXiv:2605.29766, May 2026), proposes an adaptive mixture a0=(1−w)⊙a≤t+w⊙ϵa_0=(1-\mathbf{w})\odot a_{\leq t}+\mathbf{w}\odot\epsilon. Setting w=(1−q)1\mathbf{w}=(1-q)\mathbf{1} and replacing the historical prior with BtB_t recovers your source-mixing formula Xq=qBt+(1−q)ϵX_q=qB_t+(1-q)\epsilon. Although the training objectives and flow entry mechanisms differ, the core source parameterizations substantially overlap. Could you clarify the distinctive contribution of Beyond Noise beyond adaptive prior-noise mixing?

  2. Necessity of flow entry-time coupling. Given that setting s0=0s_0=0 leaves success unchanged (Section 5.3), what evidence demonstrates that coupling entry time to qq offers a meaningful advantage over adaptive noise scaling alone?

  3. Fairness of experimental comparisons. Do all baselines receive the additional Object/Spatial demonstrations and comparable spatial conditioning? Can the improvements be isolated under matched training data and architectural components?

  4. Generality and deployment. Why does per-call refinement benefit only LIBERO-10 on clean LIBERO? What evidence supports its effectiveness under real-world conditions, given the complete absence of physical robot experiments?

Reviewer_VCqbOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how informative action priors influence flow-matching robot policies. It finds that warm starts improve performance, but much of the variation in their inputs is attenuated before reaching the executed action. Beyond Noise couples the prior weight, noise scale, and flow entry time through a single parameter, selected using validation outcomes and optionally refined by a learned regulator. Experiments on LIBERO and its perturbed variants show improvements, particularly on longer-horizon tasks.

Critical strengths
  1. The matched-source study provides useful evidence about the distinction between supplying an informative prior and preserving its influence during action generation. The perturbation experiments and source-swapping controls make this observation more concrete than a performance comparison alone.

  2. The evaluation separates validation decisions from held-out initial states and includes additional freshly sampled states. Component analyses and perturbation experiments help characterize where the method is effective and where its benefits are less consistent.

Critical weaknesses
  1. The effect of stronger prior retention on compounding errors is not directly characterized. Reusing predictions based on previous observations and actions can propagate earlier mistakes across planning calls. Increasing prior transmission could therefore preserve useful temporal information or reinforce a persistent error. The paper measures adjacent-plan consistency, local perturbation transmission, and robustness to object displacement, but these do not directly show how an initial error evolves over a long rollout. Could the authors inject a controlled error into the prior at an early planning call and track subsequent state deviation, recovery time, and regulator decisions for fixed and adaptive entry settings? This would clarify when retaining more of the prior helps recovery and when it perpetuates mistakes.

  2. The evidence for the necessity of the full coupling and dynamic regulation is mixed. Section 5.3 indicates that noise scale is the main factor controlling transmission, while entering at time zero can leave success unchanged. Moreover, the standard LIBERO evaluation enables per-call refinement only on LIBERO-10. These findings support informative low-noise initialization, but leave the incremental importance of coupling all three quantities less clear. A matched comparison against a low-noise prior with a conventional entry time and identical task-level tuning would sharpen the claimed contribution.

  3. Additional training data confound some robustness comparisons. Beyond Noise uses moved-object demonstrations for Object and Spatial. Appendix G shows that removing these demonstrations substantially reduces performance under object shift and eliminates some advantages over the baselines. This disclosure is helpful, but it limits attributing the headline robustness gains to prior regulation. Comparisons giving the same additional demonstrations to the relevant baselines would better separate the effects of data augmentation and the proposed method.

Questions for authors

See Weakness

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

Related papers

Discussion (0)

Sign in to comment.