acceptodds
Under review as a conference paper at ICLR 2027

Distinguishing Repairable Deployment Failures from Vision-Language-Action Failures Before Rollout

Abstract

Vision-language-action policies that work in the lab often fail when the camera is bumped or the arm starts somewhere new. We show that most of these failures are not the policy's fault, and that a robot can tell which ones it can fix before running the policy at all. The robot uses its own body as a measuring instrument. Its encoders say where its arm should appear in the image, and comparing that with where it does appear reveals how the camera moved. From this the robot re-renders the view the policy was trained on, predicts how much of that view it can restore, and checks whether its own measurement can be trusted. On LIBERO-Plus with three policy families, the correction repairs 61 of 100 camera failures while breaking 18 of 140 successes (p = 1.3e-6), resetting to the training start repairs all 47 starting-pose failures on untuned suites, and on a held-out policy a pre-registered test of the correction with a refined pose passes (15 repaired, 3 broken, p = 0.004). Before any rollout, the fraction of the training view the correction restores ranks which failures it will fix, even within a single perturbation category where benchmark labels carry no information, and the calibration's own residual flags every failed measurement on the held-out policies. Tracing each remaining failure to its cause, at most 8 of 100 camera failures of the development policy belong to the policy, and the rest are failed measurements or views the camera can no longer see. We prove a minimax floor on calibration from the body alone. All experiments are in simulation. With learned masks, calibration reaches about a centimetre, but the closed-loop gain over the raw policy is not yet significant, and on real images mask quality limits calibration.

Primary area:
applications to robotics, autonomy, planning
Keywords:
vision-language-action models, deployment shift, self-calibration, robustness, pre-registration

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_VFr9OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper asks whether a VLA deployment failure caused by a moved camera or a joint offset can be identified and repaired before rollout. The robot is used as its own calibration target: it moves to a few probe poses, and the camera or joint error is estimated by matching the arm silhouette. A score predicts whether the failure can be undone, and the repair is done by resetting the joints or reprojecting the image back to the training view. On LIBERO-Plus, the wrapper raises OpenVLA-OFT success under camera shifts from 0.58 to 0.76 and repairs most joint-offset cases.

Critical strengths
  1. The question is well motivated. Benchmark perturbation labels do not tell whether a failure can be fixed (e.g., a 20° orbit is recoverable while a 35° turn is not), and deciding this before rollout would be useful in deployment.
  2. Using the robot body as a calibration target to measure the nuisance and predict repairability is a good entry point, and the attempt to give a theoretical guarantee (the minimax floor on body-only calibration) is a nice idea.
Critical weaknesses
  1. The method section is hard to follow. Many components are introduced with little connection to the experiments, and it is often unclear which variant is used in which result. The theory is also not closed with the experiments.
  2. The experiments do not support the claims well. Related camera-robust VLA baselines (e.g., AnyCamVLA) are not compared. There is no sham control with the same probe motion but no reprojection. Each unit is evaluated with a single rollout, so the results are hard to reproduce. Some experiments do not show much: "repairing" all starting-pose failures by resetting to the training start pose just undoes the perturbation. In addition, one of the important meanings of the work should be repair before deployment, but no deployment in the real world is provided, which should not be a limitation.
  3. The central claim that most failures are "not the policy's fault" is largely a change of definition. A failure is attributed to the policy only if it persists under a perfect render from the training camera, i.e., the nominal condition, so this mostly restates the base success rate (as the paper itself acknowledges). Failing under a viewpoint change is itself a robustness problem of the policy. The paper uses an external wrapper to pull the input back to the training distribution rather than making the policy robust, which does not show that the policy has no deficit.
Questions for authors
  1. Can you run the sham wrapper (same probe motion, no reprojection)?
  2. How does the method compare with camera-robust VLA methods (e.g., AnyCamVLA)?
  3. Can you report results with multiple rollouts per unit?
  4. How does the theoretical bound relate to the observed repair outcomes?
  5. Why should a failure that is fixed only by restoring the training view not be considered a robustness failure of the policy?
Reviewer_wA8FOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This work proposes an external correction framework to improve the performance of frozen vision-language-action (VLA) policies under deployment distribution shifts. It estimates camera extrinsic pose using the robot’s geometric model and visual silhouette, restores the training view via depth reprojection, and resets the manipulator to its standard initial configuration via servo control.The authors also introduce a recoverability score composed of identifiability and invertibility: view invertibility is quantified by coverage, and a calibration-residual gate is integrated into the deployable version. The score ranks original policy failures by their likelihood of being rescued by correction prior to policy rollout.Simulation experiments show significant performance gains under certain settings, but the main results rely on simulator-provided segmentation and depth, with no statistically significant improvement from the learned sensing variant. While the score shows some ranking ability, the tested score-based selective correction rule fails to demonstrate a decision benefit over uniform correction.

Critical strengths
  1. The framework improves task success under several deployment-shift settings without retraining or modifying the underlying VLA policy, complementing training-based approaches to robustness.

  2. The paper evaluates multiple policy families and task suites, with ablations and analyses of calibration reliability, recoverability prediction, and failure modes. Negative results and limitations are explicitly reported, helping clarify where the approach is effective and where evidence remains limited.

Critical weaknesses
  1. The interventional ladder only confirms that correction can restore task success, but does not rule out the policy’s own sensitivity to distribution shifts. The sweeping claim that most failures are “not the policy’s fault” therefore goes beyond what the evidence supports.

  2. While the score provides predictive ranking information, the tested selective correction rules do not outperform uniform correction, as the authors themselves note. Its practical utility beyond diagnosis and failure ranking remains insufficiently established.

  3. All closed-loop evaluations are simulation-based, and core results rely on simulator-generated segmentation and depth, even with offline real-image experiments included. Real-robot experiments are needed to verify task-level benefits under realistic sensing conditions.

Questions for authors
  1. Why does restoring task success under training conditions imply that the original failure was not due to insufficient policy generalization?

  2. Since score-based selective correction does not outperform always correcting, what practical benefit does the score provide beyond prediction and analysis?

Reviewer_oV6ROpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies whether deployment-induced failures of frozen VLA policies can be anticipated and repaired before task execution. It uses the robot’s geometry and proprioception to estimate camera shifts, reconstruct the training viewpoint, and estimate repairability from calibration quality and visual coverage. The diagnostic perspective is interesting, but the distinction between deployment failures and policy robustness failures requires greater care. The methodological contribution and practical applicability also need stronger support.

Critical strengths
  1. This paper studies whether deployment-induced failures of frozen VLA policies can be anticipated and repaired before task execution. It uses the robot’s geometry and proprioception to estimate camera shifts, reconstruct the training viewpoint, and estimate repairability from calibration quality and visual coverage. The diagnostic perspective is interesting, but the distinction between deployment failures and policy robustness failures requires greater care. The methodological contribution and practical applicability also need stronger support.
  2. Using the robot’s own body as a calibration reference connects the proposed approach to measurable physical quantities. Calibration residuals and visual coverage offer understandable indicators of measurement quality and reconstruction limitations, supported by theoretical analysis under stated assumptions.
Critical weaknesses
  1. The paper acknowledges that its correction pipeline builds on existing methods, placing much of the contribution on repairability prediction and diagnosis. However, coverage alone performs comparably to the combined score, and using the prediction to gate correction does not improve the reported net repair outcome. These results leave unclear what additional practical benefit the proposed formulation provides over a simpler calibration-and-reprojection pipeline. 2.The approach requires a known reference setup, an accurate robot model, reliable robot segmentation and depth information, and safe calibration motions. Most positive closed-loop results rely on simulator sensing; the learned-sensing variant does not significantly outperform the raw policy in the reported test, and no real-robot task execution is demonstrated. Calibration motions can themselves alter subsequent performance, while the missing motion-only control makes their contribution difficult to isolate.
Questions for authors
  1. What operational definition distinguishes a deployment failure from a policy robustness failure? In particular, how would the proposed diagnosis classify a viewpoint change that preserves all task-relevant information but causes the policy to fail? Can the conclusions be separated for informative viewpoint changes and genuinely unobservable task states?
  2. What additional benefit does the proposed repairability score provide over simply applying camera calibration and viewpoint correction? Coverage alone performs similarly to the full score, and using the score to decide whether to apply correction does not improve the reported results.
  3. How effective is the method with realistic segmentation and depth estimates? Most positive results rely on simulator sensing, while the learned-sensing experiment does not show a statistically significant improvement over the original policy. What sensing accuracy is needed for the method to provide a reliable benefit?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.