acceptodds
Withdrawn from ICLR 2027

Diagnosing the Structure and Local Behavioral Effects of Viewpoint-Induced Action Distortion in Robot Policies

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Changing the external camera position can alter a visual robot policy's control commands even when the task, robot configuration, and action coordinate frame remain fixed. However, whole-episode success alone cannot separate this immediate viewpoint response from state drift accumulated after execution. To isolate the response, we query frozen policies at identical physical and policy states before execution, changing only the external-camera extrinsics. We then compare geometric, affine, information-matched linear, and conditional nonlinear mappings to characterize the resulting action differences. Bidirectional one-step action substitutions test the local behavioral effects of these differences. Our evaluation covers multiple policies, with detailed diagnosis of OpenVLA and Fast-WAM, spanning autoregressive vision-language-action and video-based world-action architectures, and complementary viewpoint evaluations on ACT and π₀. On an independent OpenVLA primary test set, a task-specific affine map raises shifted-view success from 14.0% to 49.2%, and a conditional predictor further reaches 62.4%, still below the familiar-view reference of 82.0%. At pre-contact states, injecting one shifted-view action reduces success under familiar-view continuation by 15.2 percentage points. Conversely, reverse-inserting one familiar-view action improves shifted-view continuation by 10.6 percentage points. Independent evaluation on Fast-WAM also shows one-step intervention effects and partial recovery, supporting recurrence across these two architectures. These results reveal that viewpoint changes can induce predictable action distortion before trajectories diverge, and that such changes at pre-contact interactions alter subsequent success. Our work jointly diagnoses this output structure and its behavioral effects, and provides direct evidence for locating the source of viewpoint failure and identifying the effective range of calibration.

Primary area:
applications to robotics, autonomy, planning
Keywords:
vision-language-action models, viewpoint robustness, robotic manipulation, policy diagnosis, action calibration

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

This paper studies how camera viewpoint changes affect policy actions before trajectories diverge, whether these action differences have predictable structure, and whether they affect task success. The authors keep identical physical and policy states, change only the camera extrinsics, and compare geometric, linear, and nonlinear mappings of the resulting action differences. They then use bidirectional action substitutions to test the effects on subsequent execution. Experiments on OpenVLA and Fast-WAM show that some viewpoint-induced action differences can be predicted and corrected, and that even single-step substitutions can affect task success. Additional experiments examine action components, interaction phases, and calibration transfer.

Critical strengths
  1. The paper makes a meaningful diagnostic contribution. The matched-state protocol isolates viewpoint-induced action differences before execution. This reveals effects that whole-episode success rates cannot separate from trajectory feedback.
  2. Bidirectional one-step action substitutions show that these differences can affect subsequent success. The effects recur in both OpenVLA and Fast-WAM, which is somewhat convincing.
Critical weaknesses
  1. Limited insight into physical failure mechanisms. The paper uses familiar-view actions as calibration targets, while these actions are not necessarily correct. Thus, action displacement measures disagreement with the familiar-view reference, but does not by itself establish physical control error. Although the paper visualizes action differences and example trajectories, it does not systematically show what errors the corrections resolve, such as grasp alignment, contact behavior, or gripper timing. This limits its in-depth explanation of viewpoint-induced failures.
  2. Insufficient control of confounding factors. Larger intervention effects at sampled pre-contact states do not establish that pre-contact calibration is most effective during natural shifted-view rollouts. The reason is that the phase comparisons involve different reachable subsets of source episodes. Furthermore, Task affine and Task rotation differ in gripper handling as well as continuous-action mapping. Since gripper-only correction also improves success (Figure 3(c)), their performance gap cannot be attributed solely to mapping complexity. A comparison with consistent gripper handling is needed.
Questions for authors
  1. What physical errors do the predicted corrections resolve? Can the authors quantify improvements in grasp alignment, contact behavior, or gripper timing, and provide representative before-and-after rollout visualizations?
  2. Why does OpenVLA's Native success rate at +30° differ between Table 2 (14.0%) and Table A1 (4.8%)? Beyond using independent test sets, were the initial-state sampling procedures, camera extrinsics, policy checkpoint, and rollout settings identical? If so, what explains this gap?
  3. Can the authors compare Task rotation and Task affine under identical gripper handling to isolate the effect of continuous-action mapping?
  4. Do the larger pre-contact intervention effects persist when the analysis is restricted to source episodes with eligible states in all three phases?
  5. For the primary calibration models, which frozen visual encoder provides the features used for the 64-dimensional PCA projection? How many paired calibration snapshots are used per task for training and validation?
3 comments
AuthorsOpenReview ↗

Response to Reviewer VjPC: Physical Recovery, Controlled Comparisons, and Calibration Details

We thank the reviewer for the constructive questions. Your recognition is encouraging.

Q1. Physical errors and rollout visualizations

Correction improves target acquisition and closure placement. Section 3, Eq. (2), treats familiar-view actions as references without assuming correctness. Section 5, Eqs. (6–7), separately tests their behavioral relevance.

We compared Native and the Table 2 conditional predictor on the same 500 test starts.

Physical outcomeNativeCalibrated
Stable target grasp followed by lift121/500355/500
First closure outside the feasible grasp region142/50061/500
No closure during the episode47/50032/500

Counts include failed trajectories. No-closure episodes are not counted as in-region closures. Grasp-and-lift events increase by 46.8 pp, while misplaced first closures decrease by 16.2 pp.

A separate paired example uses the supplementary CNN calibrator from Appendix D.3. From the same restored initial state, correction succeeds after 75 control steps, with bilateral target contact and a maximum target height increase of 90.7 mm. Native fails at the 220-step horizon. Event definitions and paired visualizations are provided in Appendix F.

Q2. Native success in Table 2 versus Table A1

Thank you for highlighting this discrepancy. We audited both evaluation pipelines and verified that they use the same tasks, task weights, initial-state sampling procedures, and filtering rules, with the same saved initial states used in both evaluations. The policy checkpoint, action unnormalization, and actual camera position and orientation at +30° are identical. The rollout settings also match, including the horizon, control frequency, action processing, gripper handling, and success and termination criteria.

The audit identified an image-preprocessing inconsistency. The viewpoint-sweep evaluator used for Table A1 omitted the center_crop argument when calling get_vla_action, causing it to use the default value of False, whereas Table 2 explicitly enabled center_crop=True. Thus, the reported 4.8% and 14.0% success rates were obtained under different preprocessing settings despite matching the other settings listed above. The primary Table 2 evaluation uses the intended preprocessing.

We appreciate the reviewer’s careful observation, which helped us identify this previously unreported difference between the evaluation protocols.

Q3. Rotation and affine with matched gripper handling

Affine retains higher success under matched gripper rules. We added Affine with Native gripper and Rotation with predicted gripper to the Table 2 comparison on the same 500 starts. Both predicted-gripper conditions use the same predictor and decision rule.

MappingNative gripperShared predicted gripper
Task rotation165/500190/500
Task affine226/500246/500

Affine’s gains of 12.2 and 11.2 pp support a contribution from continuous-action mapping beyond gripper correction.

Q4. Phase effects in the same source episodes

Yes. We repeated the phase comparison in Figure 3(a,b) and Table A7 using the same 110 source episodes eligible in all three phases, retaining the original phase definitions.

PhaseInjection effectRescue effect
Approach−6 (−5.5 pp)+4 (+3.6 pp)
Pre-contact−18 (−16.4 pp)+12 (+10.9 pp)
Post-contact−2 (−1.8 pp)+1 (+0.9 pp)

Entries show net changes in successful continuations. Pre-contact retains the largest point-estimated effects after matching episode membership.

In shifted-view Native roll-ins, a prospective proximity criterion selected at most one state per episode without requiring eventual contact or success, yielding 200 states from 250 episodes. This differs from the original event-relative sampling. Injection reduced familiar-view continuation success from 160/200 to 130/200 (−15.0 pp), while rescue increased shifted-view continuation success from 50/200 to 76/200 (+13.0 pp). Both actions were queried from the same restored physical and policy snapshot, changing only the current external-camera observation. Each substitution lasted one control step, with the same continuation policy within each pair.

The phase pattern persists in the common cohort, with local effects also observed after shifted-view execution.

Q5. Visual encoder and calibration sample counts

For primary OpenVLA calibration, features come from the evaluated checkpoint’s fused DINOv2–SigLIP visual backbone, kept frozen during calibration. Each of the five Spatial tasks supplies 440 training pairs and 220 validation pairs, totaling 2,200 and 1,100 paired snapshots, respectively.

Data are split by source episode, as in Table 1. As in Section 4 and Appendix B.1, PCA to 64 dimensions and normalization are fitted on training data only and fixed for validation and testing. These pairs are separate from the structural-test set.

Thank you for the rebuttal. The added controls and physical-outcome analysis address my main concerns and make the findings more convincing. I am therefore raising my rating from 3 to 4.

Thank you very much for your thoughtful review and constructive feedback. Your questions helped us strengthen the experimental controls and physical-outcome analysis. We sincerely appreciate your recognition, which means a great deal to us and motivates us to continue this line of research.

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

This paper investigates the impact of external camera viewpoint changes on the action prediction of frozen robot policies. The authors propose a pre-execution identical-state diagnostic protocol to isolate viewpoint-induced action discrepancies by restoring the same state, compare the calibration efficacy of geometric versus learning-based action mappings, and employ bidirectional single-step action replacement to evaluate their local behavioral impact. Experiments on OpenVLA and Fast-WAM demonstrate that viewpoint-induced action distortions exhibit a predictable structure, and calibrating them can significantly improve the success rate during the pre-contact phase as well as the overall manipulation success rate.

Critical strengths
  1. Tightly Controlled Action-Level Diagnosis and Local Behavioral Verification: The identical-state recovery protocol effectively decouples the immediate action shift induced by the camera from the subsequent trajectory drift. Bidirectional single-step action replacement provides direct behavioral evidence: injecting a single-step distorted-viewpoint action at a pre-contact state reduces the continuation success rate under the normal viewpoint by 15.2 percentage points, whereas inversely inserting a single-step normal-viewpoint action increases the continuation success rate under the distorted viewpoint by 10.6 percentage points.
  2. Systematic Empirical Analysis and an Insightful Recovery Baseline: The paper systematically compares geometric, affine, and information-matching linear models against nonlinear mappings, and validates them through action component and directional control experiments. Independent evaluations on OpenVLA and Fast-WAM demonstrate that a substantial proportion of viewpoint failures can be recovered via low-complexity mappings (e.g., task-specific affine transformations improve OpenVLA's success rate from 14.0% to 49.2%, while linear models raise it to 56.8%). Accompanying confidence intervals and transfer experiments also clearly delineate the effective scope and limitations of output-side calibration.
Critical weaknesses
  1. Limited Methodological Novelty and Depth of Underlying Mechanistic Explanation: Identical-state cross-viewpoint action comparisons have already been explored in prior work (e.g., Cross-View Action Consistency). While this paper effectively combines standard action mapping with action replacement experiments to demonstrate that viewpoint shifts alter output actions and impact success rates, the analysis remains confined to the controller's output interface. It does not yet reveal the underlying visual perception or internal policy representation mechanisms that cause these action distortions. Consequently, the overall contribution leans more toward solid empirical characterization rather than proposing a novel diagnostic or explanatory framework.
  2. Limited Generalizability of Core Conclusions and Inconsistent Cross-Architecture Validation Protocols: The main experiments on OpenVLA are concentrated on five highly similar LIBERO-Spatial tasks with a fixed +30-degree camera shift, relying heavily on task-specific calibration. Table 3 indicates that the transferability of this calibration to opposite viewpoints (-30 degrees, resulting in a 5.0 percentage point drop in success rate) and unseen tasks (leave-one-task-out validation yields only a 2.2 percentage point improvement) is highly limited. Furthermore, Section 7.7 of the main text switches to an alternative CNN predictor protocol detailed in the appendix. Meanwhile, ACT and pi0 are evaluated solely through result-level viewpoint recovery, without undergoing the paper's core identical-state action intervention protocol. This inconsistency restricts the generalizability of the core conclusions across diverse tasks and policy architectures.
Questions for authors
  1. Mechanistic Novelty of the Diagnostic Method: Beyond merely confirming that camera shifts alter actions and affect success rates, what specific, previously unidentified failure mechanisms does the diagnostic protocol proposed in this paper reveal, compared to existing cross-view action consistency analyses (e.g., Cross-View Action Consistency)?
  2. Actual Contribution of Visual Features in the Main Predictor: Can the authors isolate the contribution of visual features in the main conditional predictor (Table 2), i.e., by testing the closed-loop success rate while retaining only the original actions, proprioceptive states, and task identifiers? Since the "no-RGB" ablation in the supplementary experiments used a different CNN predictor and showed no performance degradation from removing direct visual input, clarifying this on the main test set would help delineate the true information sources that the calibration actually relies upon.
  3. Generalizability of Local Intervention Effects: Does the action replacement effect during the pre-contact phase still hold for tasks with different manipulation behaviors, or for states that are reached through actual execution under a shifted viewpoint (which inherently induces state drift)? Providing supplementary evidence for this would strengthen the persuasiveness of this behavioral regularity beyond the current evaluation setup.
1 comment
AuthorsOpenReview ↗

Response to Reviewer hCZA: Diagnostic Contribution and Additional Evidence

We sincerely thank the reviewer for recognizing our work and for the constructive suggestions.

Q1 and W1. Diagnostic contribution and physical outcomes

Our contribution connects predictable action structure to local behavioral consequences in frozen policies. Cross-View Action Consistency uses paired views to regularize action-flow predictions during training. Our diagnosis asks which mappings support recovery, whether correction direction matters beyond magnitude, and which physical behaviors improve with calibration.

Table 2 reports success of 14.0% for Native, 49.2% for task affine calibration, and 56.8% for the information-matched linear predictor. These results show substantial recovery through low-complexity mappings. The paired substitutions in Section 5 and Appendix B.2 test one-step action changes under the same continuation policy within each pair.

In the translation-only control, the predicted direction outperforms the mean orthogonal control by 8.3 pp (95% CI [3.3, 13.6]). Executed magnitudes are matched. Rotation and gripper commands remain Native (Figure 4(d), Appendix B.4).

We compared Native with the Table 2 conditional predictor on the same 500 test starts. Episodes with a stable target grasp followed by a lift increased from 121 to 355. Episodes with a first closure outside the feasible grasp region decreased from 142 to 61. Episodes without closure were counted separately, 47 versus 32. Failed trajectories remained in the denominator. These findings link recoverable action structure to local effects and physical improvements.

Q2. Visual features in the primary predictor

Direct visual features improve the primary predictor. We trained an ablated version of the PCA-feature MLP in Section 4 and Appendix B.1. We removed direct visual features and retained the original action, proprioception, and task identity. Both conditions used the same episode splits, matched training budgets, and 500 shared test starts.

PredictorSuccesses across three optimization seedsMean success
Full312/500, 306/500, 318/50062.4%
Without direct visual features280/500, 275/500, 285/50056.0%

The Full results appear in Appendix A.5. Removing direct visual features reduces mean success by 6.4 pp. The retained action is still produced from images by the frozen policy. This ablation therefore measures the added value of direct visual conditioning in the calibrator. It complements the separate CNN analysis in Section 7.7 and Appendix D.3.

Q3 and W2. Effects after shifted-view execution and across behaviors

Local action effects also appear after shifted-view execution and in two additional behaviors. A prospective proximity criterion selected 200 eligible states from 250 shifted-view Native episodes. Selection did not require eventual contact or success. Injection reduced familiar-view continuation success from 160/200 to 130/200 (−15.0 pp). Rescue increased shifted-view continuation success from 50/200 to 76/200 (+13.0 pp).

Both actions were queried from the same restored physical and policy snapshot. Both queries preserved the preceding shifted-view history and differed only in the current external camera observation. One control step was replaced. Within each pair, the continuation policy and the rules for updating history and action queues were identical. This prospective selection differs from the original event-relative phase sampling.

We applied the same restoration and substitution procedure using the frozen OpenVLA policy from the supplementary Goal experiments. Each task contributed 100 eligible states reached under the shifted view, selected prospectively. No calibration map was fitted or applied. Both branches used familiar-view continuation.

TaskFamiliar-view continuation baselineAfter injectionChange
Goal 0, drawer opening86/10079/100−7.0 pp
Goal 5, plate pushing92/10081/100−11.0 pp

Both tasks were excluded from primary Spatial calibration training. Both show lower point-estimated success after injection.

W2. Transfer and evidence across architectures

Table 3 reports gains without refitting of 20.0 pp at +15°, 19.0 pp at +45°, and 38.0 pp on new layouts after calibration at +30°. Success at +45° is 25.0%. Opposite-view and leave-one-task-out transfer remain limited. These results show partial transfer within the tested distributions.

Section 7.5 reports Fast-WAM injection and rescue effects of −12.0 and +12.0 pp under its matched-state protocol. ACT and π₀ provide complementary outcome-level evidence from observation-side view restoration (Figure 2(g,h), Appendix C.1). OpenVLA and Fast-WAM thus support the core diagnosis.

The additional results and detailed evaluation protocols have been incorporated into the updated supplementary material.

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

This paper investigates how camera viewpoint changes affect robot-policy actions before execution and how these changes influence subsequent behavior. Using matched physical and policy states, the authors characterize action differences through geometric, affine, and observation-conditioned mappings, then evaluate their behavioral effects through bidirectional finite action substitutions. Experiments on OpenVLA and Fast-WAM demonstrate partially predictable action distortion and meaningful pre-contact intervention effects. The paper addresses an important diagnostic question, although the underlying sources of distortion and the generality of the identified structure remain less established.

Critical strengths

Meaningful problem formulation. Separating immediate viewpoint-induced action changes from accumulated trajectory drift provides information that episode-level success rates cannot reveal.

Well-controlled behavioral evidence. State restoration, bidirectional action substitutions, and magnitude-matched direction controls strengthen the connection between action differences and downstream outcomes.

Thoughtful evaluation. Information-matched linear baselines, direct-action prediction controls, independent evaluation splits, and explicit reporting of transfer failures support a balanced interpretation.

Critical weaknesses

Limited explanation of the distortion’s source. The experiments establish behavioral consequences but do not distinguish coordinate-frame confusion from occlusion, perceptual errors, or other representation failures.

Weak evidence for transferable structure. Leave-one-task-out calibration improves success only from 14.0% to 16.2%, while opposite-view transfer yields no benefit. The learned mappings may largely capture task-specific corrections.

Calibration is not uniquely diagnostic. Direct action prediction performs similarly to residual correction. Recovery therefore also supports reference-policy approximation, without uniquely establishing a distinct distortion mechanism.

Questions for authors

Can visibility-controlled experiments distinguish geometric misalignment from missing or degraded visual information?

Do the single-step effects persist on shifted-view roll-ins and when phase comparisons use the same source episodes?

Which diagnostic findings remain consistent across unseen tasks, even when the calibration map fails to transfer?

4 comments
AuthorsOpenReview ↗

Response to Reviewer xWHB:Additional Evidence on Viewpoint Effects and Calibration Transfer

We sincerely thank the reviewer for recognizing our work and for the constructive questions.

Q1. Visibility and the source of action differences

Injection and rescue retain their estimated directions in the visibility-selected subset. Section 3 compares observations at matched physical and policy states. We assessed target and gripper visibility and projected area using simulator segmentation aligned with policy image preprocessing. Eligibility was determined in both views before intervention. Of the 200 shifted-view states in Q2, 160 met these criteria.

InterventionBaseline successAfter substitutionChange
Injection130/160108/160−13.8 pp
Rescue40/16060/160+12.5 pp

These results provide evidence that the effects are not confined to substantial target or gripper occlusion.

Geometric correction of the external camera image yielded 112/200 successes (56.0%). In a separate control under the familiar viewpoint, fully opaque masking of the critical interaction region caused a smaller reduction in success than the viewpoint shift. The image-correction results support a behavioral benefit from geometric realignment. We will conduct further mechanistic experiments to investigate whether these effects arise from coordinate-frame confusion, changes in visual representations, or their interaction.

Q2. Phase comparisons and shifted-view execution

Injection and rescue retain their estimated directions after shifted-view roll-ins. Pre-contact retains the largest phase-specific point estimates when source episodes are matched.

We restricted the comparison in Section 7.3 and Table A7 to the same 110 source episodes eligible in all three phases. We retained the original phase definitions.

PhaseInjectionRescue
Approach−5.5 pp+3.6 pp
Pre-contact−16.4 pp+10.9 pp
Post-contact−1.8 pp+0.9 pp

This addresses the phase-composition concern in Appendix B.3 within the common eligible cohort.

We also examined Native execution under the shifted view. A prospective proximity criterion selected 200 eligible states from 250 episodes without requiring eventual contact or success. Injection reduced familiar-view continuation success from 160/200 to 130/200 (−15.0 pp). Rescue increased shifted-view continuation success from 50/200 to 76/200 (+13.0 pp).

Both actions were queried from the same restored physical and policy snapshot. Both queries preserved the preceding shifted-view history and differed only in the current external camera observation. One control step was replaced. Within each pair, the continuation policy and history and queue update rules were identical, following Appendix B.2. This prospective selection differs from the original event-relative phase sampling. The results support local behavioral effects after shifted-view execution.

Q3. Recurring findings across tasks and calibration transfer

We applied the same snapshot restoration and single-step substitution procedure to drawer opening and plate pushing. We used the frozen OpenVLA policy from the supplementary Goal experiments.

Each task contributed 100 eligible states reached under the shifted view, selected using task-specific prospective criteria. No calibration map was fitted or applied. Both branches used familiar-view continuation.

TaskBaseline successAfter injectionChange
Goal 0, drawer opening86/10079/100−7.0 pp
Goal 5, plate pushing92/10081/100−11.0 pp

Both tasks were excluded from primary Spatial calibration training. Both show a lower point-estimated success rate after injection.

Table 3 separately tests transfer of a fitted map. Without refitting, calibration at +30° improves success by 20.0 pp at +15°, 19.0 pp at +45°, and 38.0 pp on new layouts. At +45°, success increases from 6.0% to 25.0%. Opposite-view and leave-one-task-out transfer remain limited. These results show recurrence of the local intervention finding in the additional behaviors and calibration gains for the tested unseen positive-angle views and new layouts.

The residual-versus-direct difference is 1.6 pp [−2.6, 5.8] (Appendix B.5). Behavioral evidence also includes finite substitutions without a learned calibrator and a separate magnitude-matched translation-direction control. The latter favors the predicted direction by 8.3 pp [3.3, 13.6] over the mean orthogonal control (Figure 3(d), Appendix B.4). Both intervals are 95% CIs.

The additional results and detailed protocols have been incorporated into the updated supplementary material.

Thanks author for the detailed rebuttal and I will keep my positive rating.

We sincerely thank you for your constructive feedback, which has helped us further improve the paper. We also appreciate your careful consideration of our response and your continued positive assessment of our work.

Dear Reviewers and Area Chair,

We provide below a summary of the revisions made to our manuscript in response to the reviewers’ comments. To incorporate the additional experiments, analyses, and visualizations, the main text has been expanded from 9 to 10 pages, within the rebuttal limit. The main changes are as follows:

  • Sections 7.2 and 7.4: We added physical-outcome measurements, an ablation of direct visual features in the primary predictor, and comparisons between rotation and affine mappings under matched gripper handling. Detailed results and protocols are provided in the newly added Appendices F.1–F.3.

  • Section 7.3: We added phase comparisons using a common source-episode cohort, single-step interventions at states reached during shifted-view execution, and visibility-based controls. These analyses and additional image-side controls are documented in the new Appendices F.4–F.6.

  • Section 7.6: We added intervention tests on two additional Goal tasks to examine local behavioral effects beyond the primary Spatial tasks. The protocols and results are provided in the new Appendix F.7.

  • Section 7.7 and new Figures 5, A12, and A13: We expanded the discussion of physical recovery and paired rollout examples. Figure 5 presents physical outcomes, matched-gripper comparisons, and selected paired rollouts. Figure A12 provides paired rollout examples across ten Spatial tasks, while Figure A13 shows recorded object height and event timing for a Task 0 example.

  • Section 4, Appendix B.1, and Section 8: We clarified the calibration implementation and the interpretation and scope of the diagnostic findings. The new Appendix F.8 documents the image-preprocessing difference between the viewpoint sweep and the primary evaluation.

We sincerely thank you for your time and constructive feedback.

Comments on the paper

2 comments
~Olivia_Brown3OpenReview ↗

Research-integrity complaint concerning ICLR 2027 Submission 56702

To the Area Chair, Senior Area Chair and Program Chairs:

I regard VistaBot (ICRA 2026) as a valuable contribution closely related to this submission, particularly in its experiments and evaluation protocol. I request a formal investigation into suspected data fabrication or falsification and misrepresentation of results in Submission 56702, and whether subsequent changes to those results have a verifiable experimental basis.

1. Earlier version: 144 identical values presented as independent evaluations

In the 33-page copy downloaded 15 Nov 2026 at 03:00 AoE, all 144 task-level success rates in Tables A9 and A10 (p. 21) exactly match Table I of VistaBot (ICRA 2026).

All 72 ACT and 72 π₀ entries match: eight tasks × (one clean reference, four default novel views and four restored views), including failure patterns and cases without restoration gains.

Yet Section 6 (lines 240–243) says “we independently evaluate,” and Figure 4 (lines 396–399) describes “Our independent evaluations of ACT and π0.” Citing VistaBot's protocol or method does not disclose numerical reuse.

Shared code does not establish identical checkpoints, samples or RNG states. Authors must document whether these were fresh samples, deterministic re-executions, reused logs or copied published values.

2. Revised version: changed results and a fourfold evaluation budget

In the 40-page copy downloaded 18 Nov 2026 at 13:00 AoE, Tables A9 and A10 (p. 22) change 32/144 values: 13 ACT and 19 π₀ entries; 112/144 (77.8%) still exactly match VistaBot. There are 18 increases and 14 decreases, including:

  • ACT, Stack wine, Default +45°: 0.00 → 0.20.
  • π₀, Stack wine, Default +30°: 0.00 → 0.20.
  • ACT, Take lid off, Default +30°: 0.32 → 0.22.

Appendix C.1 and the figure/A9/A10 captions increase 25 to 100 trials per novel-view cell, yet 102/128 (79.7%) novel-view rates remain unchanged. An unchanged rate k/25 requires exactly 4k successes in 100 trials. Episode records must explain why most rates remain identical despite four times as many trials. Six clean-reference values also change; their denominators require clarification.

Appendix C.1 changes the description from observation-side restoration to an integrated VistaBot-enhanced policy trained using view-aligned synthesis latents; whether the description or implementation changed is unexplained.

Section 6 and Figure 2 retain the independent-evaluation claim. Revised values do not explain the earlier matching table.

3. Material changes omitted from the final revision summary

The authors' final summary to the reviewers and AC describes numerous additions but does not identify the changes to Tables A8–A10, the increase from 25 to 100 trials, or the relationship between the old and new experimental records. I found no specific explanation in the supplied manuscript versions or author responses.

These are material changes to empirical evidence. They require an explicit account of what was rerun, corrected or replaced, when, and why. The chairs should also establish whether a separate explanation was provided through another channel.

4. Records and action requested

Please preserve both manuscript versions and the official revision history, and require:

  • Episode-level outcomes, episode/initial-state identifiers, seeds, checkpoint identifiers, configurations, execution records and table-generation scripts for BOTH versions.
  • A precise account of any shared data or evaluations with VistaBot, and what “independent” means.
  • Whether the later 100 trials include the original 25, replace them, or reflect a reporting correction; a mapping between trial sets where applicable.
  • An explanation of all 32 changed entries, the unchanged entries, clean-view denominators and the revised method description.
  • Appropriate uncertainty estimates and identification of independent sampling units. The tables report success-rate means and macro-averages, but not corresponding confidence intervals or variability across independent repeated evaluations.
  • An explanation of why these changes were absent from the final revision summary.

Rerun cannot establish the provenance of the results already reported. A generic appeal to identical seeds likewise cannot substitute for records linking actual evaluations to both tables.

The evidence warrants formal investigation of possible fabrication, falsification, unattributed reuse or misrepresentation; the committee should determine which findings the records substantiate. These concerns must be resolved before the submission can be assessed on its merits. Until verified, the ACT/π₀ results should not be credited as independent cross-policy validation.

I request acknowledgment and referral to the appropriate research-integrity reviewers.

Dear Area Chair, Senior Area Chair, and Program Chairs,

First and foremost, we would like to express our sincere gratitude for the time, effort, and attention you have devoted to our submission throughout the review and rebuttal process. We greatly appreciate your professionalism, constructive feedback, and careful consideration.

After careful consideration and discussion among all authors, we have collectively decided to withdraw Submission 56702 from ICLR 2027.

We would like to emphasize that the ACT and π₀ evaluations currently under discussion are supplementary cross-policy evaluations and are not the central basis of our paper's primary contributions. The core of our work lies in the matched-state diagnosis of viewpoint-induced action distortion, controlled action interventions, and the analysis of their local behavioral effects on OpenVLA and Fast-WAM.

We also wish to acknowledge that the concerns were raised publicly at a very late stage of the rebuttal process, when we had already completed our responses to the reviewers. At this stage, we are not in a position to speculate about the motivation or timing behind these concerns, nor do we wish to make assumptions about them. Our decision is instead guided by our own standards and responsibilities as researchers.

We hold our research and its scientific contributions in high regard, and precisely because we believe in the value of this work, we hold ourselves to equally high standards of experimental rigor, accuracy, and reproducibility. We do not wish to leave any unresolved questions regarding any part of our manuscript, including supplementary evaluations, regardless of how central they are to our main contributions.

Therefore, we have decided to withdraw the submission and conduct a thorough internal review of the relevant experimental records, data provenance, and evaluation procedures. This is a deliberate and precautionary decision made out of our commitment to research integrity, allowing us to verify every relevant detail to the standard that we believe our work deserves before considering any further submission or publication.

We believe that maintaining the integrity and credibility of our research is more important than proceeding with the current submission.

We sincerely thank the Area Chair, Senior Area Chair, Program Chairs, and all reviewers for their time, professionalism, and valuable feedback throughout this process. We apologize for any additional administrative effort this decision may cause and greatly appreciate your understanding.

Thank you again for your consideration.

Sincerely,
The Authors of ICLR 2027 Submission 56702

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

Related papers

Discussion (0)

Sign in to comment.