acceptodds
Withdrawn from ICLR 2027

Action-Space Velocity Correction for Visual Consistency in Robotic Flow Policies

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Visuomotor policies trained under one visual appearance degrade in performance when lighting, color, or background change during inference. Retraining or augmenting the policy for every new appearance is costly. In this paper, we introduce Action-Space Velocity Correction (ASVC), an inference-time correction to the velocity of a flow-matching policy that corrects based on a reference view of the observation. Transport conditioned on reference views provides consistency of visual variation in action space. ASVC modifies only the integration of a frozen policy, requiring no retraining and applying as an add-on to any flow-matching action policy. Our method is validated on a toy reaching task and on the DMC-GB2 benchmark across three flow-matching base policies: ReinFlow, GFP, and Diffusion Policy. We further deploy ASVC on a Trossen WidowX AI robot for a cube-stacking task with variation in nearby distractor objects and lighting. Across these settings, ASVC recovers much of the return lost under color and video shifts while performance on the training appearance is preserved.

Primary area:
applications to robotics, autonomy, planning
Keywords:
robotic policies, reinforcement learning, 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_AY1oOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Action-Space Velocity Correction (ASVC), an inference-time method that corrects Flow Matching velocity fields without retraining the original robot policy, aiming to mitigate policy performance degradation caused by visual variations in lighting, color, and background. The proposed method leverages an additionally trained reference mapping to generate observations close to the training visual distribution and selectively corrects the velocity during flow integration based on the residual projection between the original and reference velocities. The authors validate the approach on Toy Reaching, three types of base policies on DMC-GB2, and a real-world robot block-stacking task. Experimental results demonstrate that ASVC substantially recovers part of the task performance lost under visual shifts while largely preserving performance in the original environments across most settings.

Critical strengths
  1. Targeting the performance degradation caused by visual environmental changes during robot policy deployment, this paper proposes an inference-time velocity correction method without retraining the original policy. This approach reduces the cost of adapting deployed policies to new visual environments and provides a meaningful research direction for improving the visual generalization capability of robot policies.

  2. The proposed method demonstrates strong generality. ASVC can be directly integrated as a plug-and-play module into various Flow Matching policies. The authors validate its effectiveness across three distinct types of policies—ReinFlow, GFP, and Diffusion Policy—while introducing only modest additional inference overhead on real-world robot hardware.

Critical weaknesses
  1. The gating condition of Residual Projection lacks a rigorous mathematical justification. In Section 3.3, the authors project the current velocity uu onto the residual direction e^=(uclean−u)/∥uclean−u∥\hat{e} = (u_{\text{clean}} - u) / \Vert{}u_{\text{clean}} - u\Vert{} and determine whether the current velocity deviates from the reference velocity based on whether the projection value ρ\rho falls below a threshold εdir\varepsilon_{\text{dir}} (defaulting to −0.1-0.1). However, ρ<0\rho \lt 0 is mathematically equivalent to ⟨u,uclean⟩<∥u∥2\langle u, u_{\text{clean}} \rangle \lt \Vert{}u\Vert{}^2, which does not correspond to uu and ucleanu_{\text{clean}} pointing in opposite directions. For instance, when both velocities point in the exact same direction but the reference velocity has a smaller magnitude, ρ=−∥u∥\rho = -\Vert{}u\Vert{}, and the gate will still be triggered; conversely, ρ>0\rho \gt 0 does not guarantee that the two velocities share the same direction. This issue directly undermines the soundness of the core gating mechanism in ASVC and requires further clarification from the authors.

  2. Lack of a critical Reference-only baseline. ASVC relies on an additionally trained reference mapping to restore test observations close to the training visual distribution, yet the paper does not report the results of directly generating actions from the restored observations under visual shifts, and the validation of the reference mapping is conducted solely on training frames. Consequently, it is currently impossible to determine whether the performance gains stem primarily from visual restoration or from the proposed velocity correction mechanism. Notably, in the real-world robot experiment under lighting variations, the fire of ASVC reaches 100% when λ=1\lambda = 1, under which the method is strictly equivalent to integrating the velocity field corresponding to the reference observations throughout the entire process. Under shifted conditions on DMC-GB2, the fire also mostly fall within the range of 0.5–0.9, indicating that most integration steps actually use the reference velocity. Therefore, the absence of a Reference-only baseline constitutes a significant experimental flaw.

Questions for authors
  1. The gating condition of Residual Projection lacks a rigorous mathematical justification. In Section 3.3, the authors project the current velocity uu onto the residual direction e^=(uclean−u)/∥uclean−u∥\hat{e} = (u_{\text{clean}} - u) / \Vert{}u_{\text{clean}} - u\Vert{} and determine whether the current velocity deviates from the reference velocity based on whether the projection value ρ\rho falls below a threshold εdir\varepsilon_{\text{dir}} (defaulting to −0.1-0.1). However, ρ<0\rho \lt 0 is mathematically equivalent to ⟨u,uclean⟩<∥u∥2\langle u, u_{\text{clean}} \rangle \lt \Vert{}u\Vert{}^2, which does not correspond to uu and ucleanu_{\text{clean}} pointing in opposite directions. For instance, when both velocities point in the exact same direction but the reference velocity has a smaller magnitude, ρ=−∥u∥\rho = -\Vert{}u\Vert{}, and the gate will still be triggered; conversely, ρ>0\rho \gt 0 does not guarantee that the two velocities share the same direction. This issue directly undermines the soundness of the core gating mechanism in ASVC and requires further clarification from the authors.

  2. Lack of a critical Reference-only baseline. ASVC relies on an additionally trained reference mapping to restore test observations close to the training visual distribution, yet the paper does not report the results of directly generating actions from the restored observations under visual shifts, and the validation of the reference mapping is conducted solely on training frames. Consequently, it is currently impossible to determine whether the performance gains stem primarily from visual restoration or from the proposed velocity correction mechanism. Notably, in the real-world robot experiment under lighting variations, the fire of ASVC reaches 100% when λ=1\lambda = 1, under which the method is strictly equivalent to integrating the velocity field corresponding to the reference observations throughout the entire process. Under shifted conditions on DMC-GB2, the fire also mostly fall within the range of 0.5–0.9, indicating that most integration steps actually use the reference velocity. Therefore, the absence of a Reference-only baseline constitutes a significant experimental flaw.

  3. The real-world robot experiments evaluate the performance of ASVC under lighting variations and the introduction of distractors, but do not explicitly investigate changes in the workspace background. Could the authors evaluate the performance of ASVC under variations in the color and texture of the tabletop or surrounding background?

  4. The simulation experiments are primarily conducted on three DMC-GB2 tasks: walker walk, walker stand, and finger spin. Could the authors provide additional evaluations on manipulation benchmarks (such as Meta-World or Adroit) under controlled visual appearance shifts?

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

The paper proposes ASVC, an inference-time modification of a frozen flow-matching policy for robustness to visual appearance changes. A task-specific U-Net maps the observation to a reference appearance. Specifically, the policy evaluates velocities conditioned on the original and reference observations at each integration step. A gate chooses whether to correct the original velocity. The correction selects the reference-conditioned velocity with the gain used in the main experiments

Critical strengths
  • The deployment problem is useful, and keeping the action policy frozen makes the approach practically attractive. The paper describes the additional reference-map training and inference cost, rather than requiring policy fine-tuning during deployment.
  • The experiments are comprehensive. The reported improvements under visual shifts are substantial in several settings.
  • The three-seed DMC results and per-seed appendix provide useful evidence beyond selected rollout examples. Reporting correction firing rates and clean-condition degradation also helps expose the method's behavior.
Critical weaknesses
  • The experiments do not isolate the contribution of velocity gating from that of the trained reference map. The decisive baseline is the same frozen policy acting on the same mapped observation at every integration step.
  • The clean-performance claim is stronger than the evidence. The paper acknowledges it in limitations but the abstract and introduction describe preservation of training-appearance performance. The deployment trade-off should be quantified and stated consistently.
Questions for authors

Please see above. Additionally,

  • With the reference map, checkpoints, initial noise, and evaluation episodes fixed, how does ASVC compare with reference-only integration? A comparison on the existing DMC settings and robot distractor condition would directly test the gate's value. An always-blended velocity is another useful low-cost control.
  • Can you correct the geometric argument above and explain what the gate is intended to detect without identifying residual alignment with opposition between the two velocities?
Reviewer_c17tOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Action-Space Velocity Correction (ASVC), an inference-time method to improve the robustness of flow-matching robot policies under visual changes. ASVC uses a learned reference map to generate a clean observation and compares the policy's velocity predictions under the original and reference views. It selectively corrects the action-generation process without updating the policy parameters. The method is evaluated on a toy task, DMC-GB2 with three flow-based policies, and a real-world cube-stacking task. Results show improved performance under visual shifts, including lighting and background changes.

Critical strengths

The method is simple and practical. ASVC can be applied to existing flow-based policies without retraining their parameters. Correcting the action-generation process at inference time is an interesting direction for improving visual robustness.

Critical weaknesses
  1. ASVC relies on a learned reference map that restores observations toward the training appearance. However, the experiments do not directly compare ASVC against simply feeding the reference-map output to the frozen policy. This makes it difficult to determine whether the performance gains come mainly from the reference map or the proposed velocity correction. Relative ablation study can be provided to further sharpen the paper.

  2. The tasks in experiment are relatively simple. The evaluation mainly focuses on relatively simple locomotion tasks in DMC-GB2 and a single cube-stacking task on a real robot. These tasks involve relatively short-horizon behaviors and limited manipulation complexity.

Questions for authors
  1. How does ASVC compare with directly using the reference-map output as the policy observation, without velocity correction?
  2. Could you provide some experiment result to show how sensitive is ASVC to the quality of the reference map, particularly when visual changes differ substantially from the corruptions used during reference-map training?
  3. COuld you further explain why the proposed residual projection is preferable to directly measuring the directional alignment between the observed and reference velocities?

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

Related papers

Discussion (0)

Sign in to comment.