{"id":"22c60878-24fc-4c3b-968e-9dd3b9af35a2","arxiv_id":"2608.01066","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"OC-VLA++ augments camera-space action grounding with synthesized nearby views and a cross-view action-equivariance loss, improving robot manipulation success under unseen camera poses.","lead":"Robots trained with a single camera often fail when the camera moves. This paper adds a training step that generates nearby camera views and forces the robot's action predictions to agree across views, which improves performance under larger camera shifts without hurting the original view.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed viewpoint-robustness gains are within sampling noise: at the largest displacement, the Dita gap is 58/120 vs 49/120 successes (p≈0.24), and per-task differences are 1–3 rollouts.","rationale":"The reader's verdict is CONDITIONAL and already flags the absence of significance testing and small trial counts. I agree with the overall assessment, but I would elevate the statistical power issue over the geometric consistency of MoGe-2 as the most load-bearing concern. The paper's strongest claim is an empirical one: OC-VLA++ degrades more gracefully under increasing camera displacement. For that claim to hold, the observed improvements must be real effects rather than sampling noise. With 20 rollouts per task, the standard error of a success rate is about 10 percentage points, and the headline 7.5 pp gap at Position 3 corresponds to roughly 9 successes out of 120. The paper provides no error bars, p-values, or seed variance. The geometry concern is important for explaining the mechanism, but the method could in principle work even if MoGe-2 is imperfect, as long as the synthesized views provide useful training signal; conversely, if the empirical differences are noise, the central claim fails regardless of geometric correctness. I therefore recommend keeping the CONDITIONAL verdict (no change) while insisting on a significance check as a condition of acceptance. My agreement with the reader is partial because they chose the geometry assumption as the weakest link, whereas I see the lack of statistical power as the more direct threat to the central claim.","tokens_in":17669,"tokens_out":5626,"duration_ms":49479,"concrete_test":"Pool the Position 3 Dita counts into a 2x2 table (OC-VLA++ 58/120 vs OC-VLA 49/120) and run a two-proportion z-test or Fisher's exact test; if p>0.05, the headline 'slower degradation' claim is not supported. To settle the direction, rerun the same evaluation with at least 100 rollouts per task for both models and report a bootstrap 95% CI for the mean difference; the claim survives only if the CI excludes zero.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that OC-VLA++ degrades more slowly under camera displacement. The supporting evidence is statistically very weak. At Position 3, the largest shift, the Dita-based OC-VLA++ reaches 48.3% vs 40.8% for OC-VLA; across the six tasks (20 rollouts each) this is 58 vs 49 successes out of 120. A two-proportion z-test on the pooled counts gives p≈0.24, so the difference is not significant. Per-task differences are 1–2 successes (5–10 pp at n=20), which the authors themselves note: \"a 5-point difference corresponds to one additional successful rollout.\" The Qwen results show the same issue; at the fixed view the Qwen variant is actually one trial worse than OC-VLA (59.2% vs 60.0%). No confidence intervals, significance tests, or variance across seeds are reported. The simulation result in Table 2 (56.8% vs 52.4%) also lacks episode counts and significance testing; with the 100-episode protocol used in Table 4, a 4.4 pp difference would have p≈0.15. The geometry concern raised by the reader is real but secondary: even the exact-view simulation is underpowered. For the central empirical claim to hold, the measured advantages must be distinguishable from noise; the current data do not establish that.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes OC-VLA++, an extension of the OC-VLA policy that augments camera-space action grounding with geometry-guided paired-view supervision and a cross-view action-equivariance objective. For each training observation, the method uses a frozen monocular geometry estimator (MoGe-2) to synthesize nearby camera views offline, assigns each view its own camera-space action target, recovers clean action estimates from diffusion or flow-matching generative outputs, transforms both estimates into a shared robot frame, and penalizes their discrepancy. The authors evaluate on six real-robot tasks with a single training camera and on ManiSkill2, using Dita and Qwen3-VL-2B-based backbones. The central claim is that OC-VLA++ degrades more gracefully than OC-VLA as the evaluation camera moves away from the training pose, while preserving fixed-view performance.","tokens_in":17995,"tokens_out":5953,"duration_ms":49888,"significance":"The idea of explicitly teaching how camera-space action predictions should transform under viewpoint change is a sensible and potentially valuable complement to observation-centric grounding. The method is architecture-agnostic, requires no geometry or view synthesis at inference time, and the simulation setup with exact paired views is a clean way to isolate the effect. The authors should be credited for including two model families, a component ablation, and a loss-weight sensitivity analysis. If the claimed gains were statistically established, this would be a useful contribution. At present, however, the evidence does not clearly separate the reported improvements from sampling noise.","major_comments":[{"comment":"The headline claim that OC-VLA++ exhibits 'consistently slower degradation' is not statistically supported. At the largest displacement (Position 3), the Dita-based variant succeeds in 58 of 120 rollouts versus 49 of 120 for OC-VLA, a 7.5-point difference; a two-proportion z-test gives p≈0.24. The Qwen-based comparison (52 vs 45 of 120) gives p≈0.36. Per-task counts are 20 rollouts, so the reported 5–10 point per-task gains correspond to 1–2 successful rollouts, and the appendix's own statement that 'a 5-point difference corresponds to one additional successful rollout' makes the fragility explicit. No confidence intervals, error bars, or multiple-seed variance are reported, and the fixed-view Qwen comparison is actually one trial worse than OC-VLA (59.2% vs 60.0%). To support the central claim, the authors need either substantially more rollouts, seed variance, or a pre-specified significance test on the degradation rate.","section":"Camera-Displacement Evaluation, Figure 4 and Table 6"},{"comment":"The simulation results are similarly underpowered. Table 2 reports 56.8% versus 52.4% for OC-VLA++ versus OC-VLA; under the 100-episode-per-task protocol stated for Table 4, this is 284 vs 262 successes out of 500, giving p≈0.16. The ablations in Table 3 show margins of 1.0, 1.8, and 2.6 percentage points over the OC-VLA baseline, each consistent with chance at these counts. If the simulation is intended to provide a controlled setting with exact paired views, it is exactly the place where the effect should be visible at high power; currently it is not. Please report episode counts for Table 2, standard errors, and at least one independent seed per configuration.","section":"Simulation Experiments, Tables 2 and 3"},{"comment":"The correctness of the cross-view equivariance target depends on the metric accuracy of MoGe-2's point maps. If the reconstructed geometry is biased or miscalibrated, the synthesized-view camera-space action target A^{c'}_t computed via Eq. (3) will be wrong, and the equivariance loss (Eq. 11) will align predictions to erroneous robot-frame actions. The paper currently supports this premise with a binary valid-pixel filter and visual examples (Figure 8), which do not quantify geometric error. I suggest adding a quantitative validation of the synthesized-view action targets (e.g., reprojection or action-target error against known camera poses, or a simulation experiment with deliberately perturbed geometry) to bound the sensitivity of the method to monocular reconstruction error.","section":"Eq. (2) and Paired-View Generation Details"}],"minor_comments":[{"comment":"The abstract contains a typo: 'We proposeOC-VLA++' is missing a space between 'propose' and 'OC-VLA++'.","section":"Abstract"},{"comment":"The displacement curves would be much more informative with confidence intervals or per-point standard errors, given the small per-position sample sizes.","section":"Figure 4"},{"comment":"The abbreviations 'Synth.' and 'Equiv.' are not defined in the caption; please spell out 'synthesized-view generative objective' and 'cross-view action-equivariance loss' in the caption itself.","section":"Table 3 caption"},{"comment":"The caveat that 'a 5-point difference corresponds to one additional successful rollout' is placed only in the appendix; this limitation is important enough to be stated in the main text alongside the headline numbers.","section":"Appendix, Detailed Results under Camera Displacement"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's main obstacle is statistical underpowering: the headline differences are within sampling noise, and the absence of significance testing makes the central claim difficult to accept as stated. The paper also relies heavily on the authors' own OC-VLA baseline and evaluation protocol, and it would be stronger with an independent reproduction or at least per-seed results. The geometry-validation concern is real but secondary; the immediate priority should be establishing that the measured advantages are not noise."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First, the takeaway: the method is a sensible extension of OC-VLA, but the current evidence does not show the claimed viewpoint-robustness gain. The central comparison at the largest camera shift—48.3% vs 40.8% for Dita—is 58 vs 49 successes out of 120 trials, which a two-proportion test puts around p≈0.24. Per task, a 5-point gap is one successful rollout. The paper itself acknowledges this in the appendix. So the headline claim of \"substantial improvements\" is not supported.\n\nWhat is new: the explicit cross-view action-equivariance loss, where paired views are assigned view-specific camera-space targets and regularized in the shared robot frame. That is a different mechanism from image-level augmentation or view synthesis. The presentation is clear; the equations are laid out carefully, and the implementation details—Huber losses, discrepancy caps, stop-gradient on the reference branch—are thoughtful. The fact that fixed-view performance is unchanged is a fair negative result, and the ablation on ManiSkill2 has the right structure.\n\nThe soft spots are mostly empirical. There are no error bars, no seed variation, no significance tests anywhere. The ManiSkill2 result (56.8 vs 52.4 over five tasks) has the same problem; 4.4 points over 100 or 500 episodes is within noise. The geometry assumption is real but secondary: if MoGe-2's point maps are off, the equivariance target is wrong. The visual examples look reasonable, but there is no quantitative check of reprojection error or action-target correctness. And the paper compares only against OC-VLA and a robot-base baseline; no view-synthesis or camera-conditioning method from the related work is included, which makes it hard to know how this compares to existing viewpoint-robust approaches.\n\nThe trend across tasks, models, and camera positions is consistent, so I wouldn't call the idea wrong. But with n=20 per task, this is a pilot study. The fix is not deep: more rollouts per condition, confidence intervals, and at least one established baseline would make the claim testable. For the robot-manipulation community, the idea is worth engaging with. I would send it to review, but with a clear request for stronger evidence before publication.","headline":"The cross-view equivariance idea is genuinely new and worth airing, but the reported gains sit within sampling noise, so the paper needs stronger evidence before its claims can be taken at face value.","tokens_in":18500,"tokens_out":2297,"would_cite":false,"duration_ms":21022,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"OC-VLA++ claims that explicitly teaching a robot policy how its predicted actions must transform when the camera moves—through synthetic paired views and a cross-view action-equivariance loss—makes it more robust to unseen camera poses…","keywords":["vision-language-action model","viewpoint robustness","cross-view action equivariance","monocular geometry","paired-view supervision","camera-space action grounding","robotic manipulation","diffusion policy"],"falsifier":"Take a real manipulation scene, capture RGB-D from two calibrated cameras, and train OC-VLA++ with paired views synthesized from the monocular point map; then compare each synthesized view's camera-space action target, computed from the known transform, against the action obtained by transforming the true robot action through the calibrated extrinsics. If the discrepancy exceeds the loss capping thresholds on a large fraction of pixels, the equivariance loss is aligning predictions to erroneous targets, and a variant trained with exact rendered pairs, as in simulation, should outperform it on the same real-robot displacement protocol.","tokens_in":17474,"feed_emoji":"🤖","tokens_out":5971,"duration_ms":47736,"temperature":0.7,"pith_summary":"This paper argues that the remaining weakness of camera-space action grounding is not the representation itself but the lack of cross-view supervision: a policy trained from one or a few fixed cameras has never been taught how its action predictions should change when the camera moves. OC-VLA++ therefore pairs every training observation with a nearby synthetic view, generated offline by reprojecting a monocular metric point map through a sampled camera perturbation, and gives each view its own camera-space action target. The two branches share parameters, and the recovered clean-action estimates are transformed into a shared robot frame and aligned by a cross-view action-equivariance loss. On a real robot with 20 rollouts per task, the largest camera shift sees average success of 48.3% with the Dita-based policy and 43.3% with the Qwen3-VL-2B-based policy, versus 40.8% and 37.5% for OC-VLA, while fixed-view success stays at 68.3% and 59.2%.","feed_headline":"Robot policies keep working when the camera moves","feed_subtitle":"Paired-view action equivariance preserves fixed-view skill while boosting unseen-view success by up to 7.5 points.","key_machinery":"The load-bearing object is the paired-view training signal built from monocular geometry. For each real observation, a frozen monocular geometry estimator produces a metric point map and validity mask; local camera perturbations within 15 cm translation and 10 degrees rotation are sampled, the point map is rigidly transformed and reprojected to synthesize a nearby view, and the known camera transform gives that view's camera-space action target. The original and synthesized views share model parameters and the same generative timestep, each is supervised by its own generative loss, and a differentiable clean-action estimate is recovered from each branch and transformed into the robot frame. The cross-view action-equivariance loss then compares translation, quaternion geodesic rotation discrepancy, and gripper command between the synthesized-view prediction and a stop-gradient reference from the original view, with capped Huber terms to limit the influence of synthesis artifacts. This mechanism explicitly teaches the model how actions must transform when the camera moves.","core_discovery":"The central claim is that viewpoint generalization in robot manipulation can be improved by supervising how actions transform across viewpoints directly, rather than by relying on image-level augmentation or canonicalizing observations to a reference view. Starting from OC-VLA's camera-space action representation, the paper adds two complementary training signals: generative supervision on a synthesized nearby view with that view's correct camera-space target, and an equivariance loss that aligns recovered robot-frame actions from paired views. The design is architecture-agnostic, applying to both diffusion denoisers and flow-matching action experts, adds no new network modules, and is used only at training time. The paper supports the claim with real-robot experiments under increasing camera displacement, simulation experiments with exact paired views, and ablations showing that each training component contributes separately.","pith_inferences":["If the equivariance premise holds, the same paired-view objective could be applied to any camera-centric action representation, including ones that predict relative to the gripper or an object frame, where synthesizing views would be even cheaper.","The valid-pixel filter and capped loss terms imply the method should degrade gracefully as monocular geometry quality worsens; a direct stress test would be to perturb the point map accuracy and measure how much of the gain survives.","The finding that local perturbations improve large-displacement robustness suggests that sampling perturbations from a broader distribution, or an online curriculum over perturbation size, may yield further gains without new architecture.","The equivariance loss could be reused at test time as a self-check: disagreement between real and synthesized views about the same robot-frame action could flag viewpoint failures before execution, which the paper does not explore."],"forward_implications":["Policies trained from a single fixed camera can be deployed at camera positions tens of centimeters and tens of degrees away, with a smaller success-rate drop than the camera-space baseline.","The training recipe carries across generative model families: both a diffusion-transformer policy and a flow-matching action-expert policy benefit, so the method is not tied to one action head.","Fixed-view performance is essentially unchanged, so the robustness gain does not trade away in-distribution accuracy.","Local paired-view supervision within 15 cm and 10 degrees transfers to larger evaluation shifts up to about 30.5 cm and 23.4 degrees, suggesting the mechanism generalizes beyond the sampled neighborhood.","At inference time no geometry estimator or synthesis branch is needed, so deployment cost is identical to the base policy."],"supporting_citations":[{"why":"Defines OC-VLA, the observation-centric camera-space action grounding baseline that OC-VLA++ extends and directly compares against.","marker":"(Zhang et al. 2026)"},{"why":"Supplies the Dita diffusion-transformer architecture used as one of the two model instantiations.","marker":"(Hou et al. 2025)"},{"why":"Supplies the Qwen3-VL-2B vision-language backbone used in the flow-matching action-expert instantiation.","marker":"(Bai et al. 2025)"},{"why":"Supplies the action-expert formulation and flow-matching objective that the Qwen-based variant follows.","marker":"(Black et al. 2024)"},{"why":"Supplies the frozen MoGe-2 monocular geometry model that produces the metric point maps used to synthesize paired views.","marker":"(Wang et al. 2026b)"},{"why":"Supplies the ManiSkill2 simulation benchmark and protocol used for controlled paired-view evaluation and ablations.","marker":"(Gu et al. 2023)"},{"why":"Supplies the DROID pretraining data used to initialize the Dita-based policy before target fine-tuning.","marker":"(Khazatsky et al. 2024)"}],"fun_headline_variants":["Cross-view action equivariance keeps robot skills camera-proof","Robots learn to act consistently no matter where the camera sits","Geometry-guided paired views boost robot viewpoint generalization","Camera displacement? Robot manipulation still works via view equivariance","Teaching robots cross-view consistency for robust task execution"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The synthesized paired views, produced by reprojecting a monocular point map under a sampled camera transform, are geometrically accurate enough that the camera-space action target derived from the known transform is correct supervision; if the monocular metric geometry is off, the equivariance loss can train the model toward actions that are right for a view that does not exist.","fun_headline_variants_meta":{"raw":{"variants":["Cross-view action equivariance keeps robot skills camera-proof","Robots learn to act consistently no matter where the camera sits","Geometry-guided paired views boost robot viewpoint generalization","Camera displacement? Robot manipulation still works via view equivariance","Teaching robots cross-view consistency for robust task execution"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000668,"raw_usage":{"total_tokens":3003,"prompt_tokens":858,"completion_tokens":2145,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":474,"completion_tokens_details":{"reasoning_tokens":2069}},"tokens_in":474,"tokens_out":2145,"duration_ms":15071,"temperature":1.0,"reasoning_tokens":2069,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T15:13:36.189601+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a real manipulation scene, capture RGB-D from two calibrated cameras, and train OC-VLA++ with paired views synthesized from the monocular point map; then compare each synthesized view's camera-space action target, computed from the known transform, against the action obtained by transforming the true robot action through the calibrated extrinsics. If the discrepancy exceeds the loss capping thresholds on a large fraction of pixels, the equivariance loss is aligning predictions to erroneous targets, and a variant trained with exact rendered pairs, as in simulation, should outperform it on the same real-robot displacement protocol.","supporting_citations":[],"review_version":1}