{"id":"833f1197-2985-45d7-b1d3-7374da370437","arxiv_id":"2508.13103","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":4.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"OC-VLA re-labels robot action targets from the robot base frame to the camera frame using the camera's extrinsic calibration, improving cross-view generalization of VLA policies.","lead":"This paper proposes a simple change to how robots learn from video: express the robot's hand movements in the camera's own coordinate frame rather than the robot's body frame. If the change works as claimed, it is a low-cost way to make robot control models more reliable when camera positions change.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Camera-frame action labels alone do not create camera-invariant representations; the cross-view-gain claim needs an ablative test that isolates relabeling from retraining on the new view.","rationale":"The reader's verdict is UNVERDICTED because the full text is unreadable; I agree that the evidence base is insufficient to ACCEPT, and no new information in the corrupted body changes that. The reader's weakest assumption (accurate, stable extrinsics) is real, but my reading of the abstract suggests a more fundamental concern: the mechanism itself is asserted, not derived. A pure label re-parameterization is a bijective change of coordinates applied to the regression target; it cannot add information about camera geometry that was not already present in the image. For it to improve cross-view generalization, the benefit must come from inductive bias (spatial alignment of action targets with image features) and/or optimization conditioning. That inductive-bias story is plausible and is exactly the kind of thing that COULD improve generalization, so this is not an internal-inconsistency objection. It is a correctness-risk objection about the empirical support: the abstract claims convergence speed, higher success rate, and cross-view generalization without reporting the necessary ablations. The stress-test check I propose would settle whether relabeling is the active ingredient. Given the damaged full text, agreeing with UNVERDICTED is the honest outcome, but the concrete concern is sharper than the reader's extrinsic-calibration concern.","tokens_in":18180,"tokens_out":1571,"duration_ms":15443,"concrete_test":"Run the reported real-robot or simulation benchmark in four configurations matched for total optimizer steps and data: (1) trained on camera A only, base-frame labels; (2) trained on cameras A+B, base-frame labels; (3) trained on cameras A+B, camera-frame labels (OC-VLA); (4) trained on camera A with camera-frame labels, deployed on camera B with exact extrinsics recomputed. Compare success rates and convergence curves. If (3) performs no better than (2), convergence gains come from extra data, not unification. If (4) fails dramatically, the cross-view claim depends on accurate deployment-time extrinsics. Additionally, perturb the deployment extrinsics by 1 cm / 1 deg rotation and report success-rate drop; this directly tests robustness of the camera-frame prediction target.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Central claim: transforming end-effector targets from base frame to camera frame via extrinsic calibration unifies prediction targets across viewpoints, accelerating convergence and improving cross-view generalization. The load-bearing condition is that relabeling alone confers the benefit. This is insecure for two reasons. (1) The label transform does not change the representation: the pixel encoder is unchanged, so the network must still infer the observing camera's identity/pose implicitly from the image. Camera-frame targets are not a camera-invariant representation; they are a camera-dependent representation aligned with the current view, and the transform's main effect is to remove the need for the policy head to learn the per-camera base-to-camera mapping. (2) The paper does not show as a control a baseline that trains on the new camera view with base-frame targets and the same number of steps/data; reported gains in convergence and success rate could come from additional optimization steps, from normalization/unit-scale benefits (camera-frame target deltas are typically smaller and better conditioned than base-frame poses), or from training on a single camera rather than multiple cameras. Without ablating (a) base-frame targets given identical multi-view training, (b) per-view heads, and (c) camera-frame targets using stale/noisy extrinsics, the headline mechanism is not established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes OC-VLA, a method that relabels end-effector action targets from the robot base coordinate frame into the camera coordinate frame of the observation camera, using the camera's extrinsic calibration matrix. The authors argue that this 'observation-centric' relabeling aligns action prediction with the pixel observation and thereby unifies prediction targets across heterogeneous viewpoints. They claim that OC-VLA accelerates convergence, improves task success rates, and improves cross-view generalization, and that it is a lightweight, plug-and-play modification compatible with existing VLA architectures. The paper reports evaluations on simulated and real-world robotic manipulation tasks.","tokens_in":18402,"tokens_out":4719,"duration_ms":54397,"significance":"If the empirical claims hold, the contribution is a simple and practical training-time relabeling that could improve the viewpoint robustness of VLA policies. The coordinate transform itself is mathematically straightforward and unlikely to be incorrect, and the method does not introduce free parameters fitted to the evaluation data, so the circularity burden is low. The main value of the paper would be the empirical demonstration of a causal benefit from the relabeling, and that demonstration is currently missing from the manuscript. The paper would be publishable if it provides clean ablations that isolate the relabeling mechanism from confounding factors such as target normalization, training budget, and data mixture.","major_comments":[{"comment":"The central claim that camera-frame targets improve cross-view generalization is not supported by an ablative test that isolates relabeling from other factors. The experiments do not appear to include a baseline trained on the same multi-view dataset with base-frame targets, matched in architecture, number of steps, batch size, and data augmentation. Without such a control, the reported gains in convergence and success rate in Tables 1–3 could be caused by better-conditioned target scales, additional optimization steps, or the use of a single camera instead of multiple cameras during training. Please add this baseline and report per-seed mean and standard deviation.","section":"§4 (Experiments)"},{"comment":"The paper describes the method as 'unifying prediction targets across heterogeneous viewpoints,' but camera-frame targets are not camera-invariant: for a fixed physical end-effector pose, the numerical label changes when the camera pose changes. The model must still infer the observer's viewpoint from the image. The actual contribution is a relabeling strategy that changes the target coordinate frame, not a unified or view-invariant representation. This overstatement matters because it leads to an evaluation design that may not test true cross-view generalization; the cross-view experiment should use a held-out camera pose never seen during training, rather than a held-out scene with the same training viewpoints.","section":"§3 (coordinate transform)"},{"comment":"The method relies on accurate extrinsic calibration for every camera used in training and deployment, but the manuscript does not analyze sensitivity to calibration error. In real-world robotic setups, camera mounts can shift or drift after calibration. A perturbation study that adds, for example, 1–5 cm translation error and 1–5 degrees rotation error to the extrinsics used to relabel the evaluation targets would be needed to determine whether the claimed robustness to viewpoint variations is robust to calibration noise. Without this, the practical scope of the claim is unclear.","section":"§4 (calibration robustness)"},{"comment":"The reported success rates are not accompanied by error bars, number of seeds, or statistical significance tests. Because the abstract claims substantial improvements, the lack of variance information makes it impossible to assess whether the differences are meaningful. Please report means over multiple seeds with standard deviations, and state the number of evaluation episodes per task.","section":"§4 (Tables 2–3)"}],"minor_comments":[{"comment":"The name 'Observation-Centric VLA' suggests that the observation representation is changed, but the method actually changes the action label frame. Consider renaming the method or clarifying in the introduction that the relabeling is in the action space, not in the representation.","section":"Abstract"},{"comment":"The coordinate transform equations should define all symbols consistently, for example T_camera^base and p_ee^camera = T_camera^base * p_ee^base, and should specify whether orientation is represented as Euler angles, quaternions, or rotation matrices, because the choice affects target continuity and the scale of the regression loss.","section":"§3 (Method)"},{"comment":"The related work discussion should cite recent VLA systems such as RT-2, OpenVLA, and π0, and should compare against prior work on action-space normalization, canonicalization, or frame-invariant action representations.","section":"§2 (Related Work)"},{"comment":"The tables would be easier to read if each task name were spelled out and if the baseline rows were explicitly labeled with the action-frame convention (base frame vs. camera frame) and the number of training views.","section":"§4 (Tables)"}],"recommendation":"major_revision","confidential_remarks":"The scientific idea is plausible and the coordinate transform is simple, but the empirical evidence currently does not establish the headline claim because the experiments lack the necessary ablations. I could not fully verify the experimental details from the submitted text, so a clean resubmission with complete tables, baselines, and calibration-robustness results is needed before a final accept/reject decision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Take the abstract at face value: the one real idea here is to relabel end-effector targets from the robot base frame to the camera frame using extrinsics, so the policy head regresses poses in the observing camera's coordinates. That is a legitimate, cheap, and easy-to-adopt trick. It is not a new framework and not deep theory; the transform is a standard rigid-body change of coordinates. The credible contribution, if the experiments hold up, is an engineering fix: a VLA trained this way may converge faster and generalize better across views because the head no longer has to learn separate base-to-camera maps for every camera.\n\nThe abstract's claims are plausible but currently uncheckable. The full text we have is corrupted beyond reading, so I'm working from the abstract alone. I would trust the transform math; I would not trust the empirical claims until I see the ablations. The stress-test note has the right worry: labeling in camera frame does not make the representation camera-invariant. The pixel encoder still sees a camera-dependent image, and the network must still reason about which camera it is looking at. What the relabeling can do is remove part of the policy head's burden and give better-conditioned targets. That is a real benefit, but it is not the same as making the policy robust to viewpoint changes. To justify the cross-view claim, the paper needs an ablation that trains on a new camera view with base-frame targets and the same number of steps and data, and shows a gap. It also needs a sensitivity test for stale or noisy extrinsics, because the whole trick depends on calibration being right. The abstract provides none of that.\n\nThe circularity burden is low, no hidden fitted parameters, and the idea is clearly presented. This is a modest within-subfield improvement, not a field reorganisation. The citation pattern visible in the abstract is unremarkable.\n\nWho it is for: people training VLAs on real or simulated robot data with multiple cameras. If you work in that area, this is a ten-minute read worth a citation if the ablations are honest.\n\nRecommendation: send it to peer review, but tell the editor this one needs a strict referee. The referee should ask for the base-frame same-data ablation, an extrinsic noise sweep, and a table with standard errors. If those are missing, major revision, not accept. I would not desk-reject; the idea is too simple and too plausible to ignore.","headline":"A simple, plausible coordinate-relabeling trick with an overstated cross-view-generalization claim that needs the ablations it doesn't show.","tokens_in":18906,"tokens_out":2562,"would_cite":false,"duration_ms":26788,"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":"Grounding action labels in the observing camera's coordinate frame makes vision-language-action policies converge faster and generalize across viewpoints.","keywords":["vision-language-action models","robot manipulation","camera coordinate frame","extrinsic calibration","cross-view generalization","end-effector pose prediction","observation-centric action grounding"],"falsifier":"Train a VLA with camera-frame targets on data from several cameras, then evaluate it at a held-out viewpoint never seen in training and compare against the same policy trained with base-frame targets: if the camera-frame policy does not clearly outperform the base-frame baseline, the cross-view generalization claim is falsified. A second check is to perturb the extrinsic matrices during training or deployment by a few centimeters and degrees; under the paper's logic, this should corrupt the camera-frame labels and degrade OC-VLA more than a base-frame policy.","tokens_in":1192,"feed_emoji":"🤖","tokens_out":3190,"duration_ms":66382,"temperature":0.7,"pith_summary":"Vision-language-action policies usually predict end-effector poses in the robot base frame, but training data comes from many cameras, so the same action gets different labels depending on which camera saw it. This paper proposes labeling each action in the coordinate frame of the camera that observes the scene, using the camera's extrinsic calibration matrix to transform base-frame poses into camera-frame poses. The authors argue this unifies the prediction targets across heterogeneous viewpoints, aligning actions with the pixels the policy sees. They report that this lightweight, plug-and-play change accelerates convergence, raises success rates, and improves robustness to camera viewpoint changes on both simulated and real manipulation tasks. If correct, it would make VLA policies more reliable when cameras move or are swapped.","feed_headline":"Camera-space actions make robot policies generalize across views","feed_subtitle":"A calibration-based transform unifies action targets across cameras, so policies converge faster and fail less often.","key_machinery":"The central object is the camera extrinsic calibration matrix: a rigid rotation-and-translation transform that maps a point from the robot base frame into the camera frame. OC-VLA applies this matrix to every end-effector pose label before training, so the policy's regression targets share the same coordinate system as its visual observations. This single deterministic transform does the work, which is why the method is described as lightweight and plug-and-play.","core_discovery":"The central claim is that action targets should be expressed in the observing camera's coordinate system rather than in the robot base frame. For each training episode, the end-effector pose is transformed from base coordinates to camera coordinates using the camera's extrinsic calibration matrix. This makes labels from different cameras comparable and aligned with the visual input, so the policy learns a more consistent mapping from pixels to actions. The method, named Observation-Centric VLA (OC-VLA), requires no new network modules or auxiliary losses; it is applied as a preprocessing step and is therefore compatible with existing VLA architectures. The paper reports that this coordinate change alone accelerates convergence, improves task success, and strengthens cross-view generalization in simulations and real-world experiments.","pith_inferences":["If the central claim holds, viewpoint augmentation and domain randomization might become less necessary, because the label distribution is already invariant to the base frame; the paper does not explicitly draw this conclusion.","The benefit should be small when all training data comes from a single fixed camera, since base-frame and camera-frame targets differ only by a fixed rigid transform; this makes the claim testable by varying camera diversity.","Extending the method to eye-in-hand cameras would require online extrinsic updates, since the camera moves with the arm; the static-calibration assumption would need revisiting in that setting."],"forward_implications":["A VLA trained with camera-frame action targets should reach a given success level with fewer training steps than the same architecture trained on base-frame targets.","Policies trained this way should retain success when the deployment camera is moved, re-oriented, or replaced.","The transform can be added as an input preprocessing step to existing VLA pipelines without architectural changes.","Multi-camera training data no longer carries viewpoint-dependent label inconsistencies, so the policy sees a more coherent action distribution."],"supporting_citations":[],"fun_headline_variants":["Ground robot actions in camera space for robust generalization","Camera-space actions yield more generalizable robot policies","OC-VLA: ground actions in the camera's frame","Align action labels to the camera view for stronger VLA","Robots act in camera space to generalize across views"],"cache_read_input_tokens":21120,"weakest_assumption_plain":"Every camera used in training and deployment must have an accurate, stable extrinsic calibration matrix; if calibration is noisy, stale, or missing, the camera-frame targets are misaligned with the pixels and the claimed gains vanish.","fun_headline_variants_meta":{"raw":{"variants":["Ground robot actions in camera space for robust generalization","Camera-space actions yield more generalizable robot policies","OC-VLA: ground actions in the camera's frame","Align action labels to the camera view for stronger VLA","Robots act in camera space to generalize across views"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000616,"raw_usage":{"total_tokens":2818,"prompt_tokens":860,"completion_tokens":1958,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":476,"completion_tokens_details":{"reasoning_tokens":1882}},"tokens_in":476,"tokens_out":1958,"duration_ms":13657,"temperature":1.0,"reasoning_tokens":1882,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:16:19.216181+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a VLA with camera-frame targets on data from several cameras, then evaluate it at a held-out viewpoint never seen in training and compare against the same policy trained with base-frame targets: if the camera-frame policy does not clearly outperform the base-frame baseline, the cross-view generalization claim is falsified. A second check is to perturb the extrinsic matrices during training or deployment by a few centimeters and degrees; under the paper's logic, this should corrupt the camera-frame labels and degrade OC-VLA more than a base-frame policy.","supporting_citations":[],"review_version":2}