- 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
- 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.
- 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
- 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.
- 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
- 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?
- 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?
- Can the authors compare Task rotation and Task affine under identical gripper handling to isolate the effect of continuous-action mapping?
- Do the larger pre-contact intervention effects persist when the analysis is restricted to source episodes with eligible states in all three phases?
- 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
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.
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.
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_cropargument when callingget_vla_action, causing it to use the default value ofFalse, whereas Table 2 explicitly enabledcenter_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.
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.
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.