{"id":"8767a5f4-d70b-4747-a527-114afa6b6293","arxiv_id":"2510.03011","paper_version":2,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A diffusion policy generates ordered spray-painting trajectories from point clouds, but its claimed coverage advantage reverses against the paper's own strongest baseline on three of four categories.","lead":"A robotics paper trains a diffusion model to generate spray-painting robot paths from 3D point clouds, claiming large gains in coverage and smoothness over earlier methods. The paper's own tables show its surface coverage is below a simpler baseline on three of four object categories, and the headline improvement numbers do not match its reported results.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Abstract's headline margins are not reproducible from Table I: coverage vs Point-Wise averages −11.4 pp, not +67.5 pp; PCD reduction vs PaintNet averages 93.2%, not 98.2%.","rationale":"The reader's weakest_assumption focuses on the unmatched trajectory-history conditioning channel (Sec. III.C.a, Eqs. 3-4; Table III), which is a serious confound for the PCD and jerk comparisons: removing it raises PCD from 10.41 to 246.07 on Windows. I agree that this undermines the comparative claims. However, I believe the more load-bearing concern is that the abstract's headline numbers are not reproducible from the paper's own Table I, and this does not depend on any assumption about baseline fairness. The coverage metric—described as the primary evaluation metric—is lower than the Point-Wise baseline on three of four categories, and the average deficit is 11.4 pp. The abstract's '+67.5 pp' (or '+61%') is therefore directly contradicted by the body. This is a mechanical falsification of the central claim, not a matter of interpretation. The reader's rationale already notes this (points 1-2), so our disagreement is only about which concern is primary. The concrete check is a simple arithmetic audit, which should settle it. The paper has useful components—the 3DCovDiff encoder ablation and the trajectory-conditioning study are informative—but the central comparative claim as stated cannot stand.","tokens_in":12548,"tokens_out":7210,"duration_ms":64821,"concrete_test":"Perform a direct arithmetic audit of Table I: compute the per-category relative improvements for PCD and jerk and the percentage-point differences for coverage, for each baseline (PaintNet, Multi-Path, Point-Wise), and average over categories. Confirm whether any consistent baseline-selection rule produces the abstract's figures (98.2% PCD, 97.0% jerk, +67.5 pp or +61% coverage). No new experiments are needed; if the advertised numbers are not reproducible, the abstract's central comparative claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table I is the only quantitative comparison behind the abstract's central claim. Recomputing the aggregate margins from Table I does not reproduce any of the three advertised numbers. PCD reduction vs PaintNet averages (98.5+99.3+98.8+76.3)/4 = 93.2%, not 98.2%; vs Point-Wise, the Containers category is 2x worse (622.16 vs 313.98), so no average across all four categories can reach 98.2%. Jerk reduction vs PaintNet averages 85.7%, and only reaches ~97.8% when compared selectively to the weakest baseline (Point-Wise). Coverage: vs PaintNet the average gain is +58.2 pp, not +67.5 pp; vs Point-Wise the average change is −11.4 pp, with Ours lower on Cuboids, Shelves, and Containers (Table I). The paper's Section IV.A.b acknowledges only the Windows gain over Point-Wise and omits these losses. Since coverage is called the 'primary evaluation metric' in Section IV, the body's own numbers contradict the abstract's '+67.5 percentage points' and 'increasing surface coverage by 61%.' This is an internal inconsistency, independent of the trajectory-history conditioning confound (Sec. III.C.a, Table III), and it falsifies the central claim of 'substantial gains over prior learning-based baselines.'","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes 3D-CovDiffusion, a DDIM-based conditional diffusion policy that generates ordered 6-DoF spray-painting trajectory chunks from a partial point cloud and a 4-step motion-history condition. The authors claim that a single policy, without category-specific training, can generalize across cuboids, windows, shelves, and containers, and that it outperforms PaintNet, Multi-Path Prediction, and Point-Wise Prediction by large margins in point-wise Chamfer distance, surface coverage, and jerk-based smoothness. The evaluation uses an extended PaintNet/MaskPlanner dataset with 80/20 splits. The paper includes ablations on the point-cloud encoder and on the trajectory-conditioning mechanism.","tokens_in":12759,"tokens_out":10232,"duration_ms":80157,"significance":"The problem is relevant: replacing heuristic segment stitching with a learned, geometry-conditioned sequence generator would be a practical advance for industrial surface processing. If the claimed gains were real, the work would provide a useful baseline for diffusion-based coverage path planning. The paper also has some positive features: it reports mean±std over seeds, runs ablations of encoder choice and trajectory conditioning, and provides qualitative comparisons. However, the quantitative claims are not supported by the paper's own table, and the main comparison is confounded by an input channel (the 4-step history) that baselines do not receive. These are internal problems requiring substantive rework, not presentation fixes.","major_comments":[{"comment":"The abstract's headline numbers are not derivable from Table I. Coverage: the average gain over PaintNet is +58.2 pp, not +67.5 pp, and vs Point-Wise it is -11.4 pp (Ours is lower on Cuboids, Shelves, Containers). PCD: the average reduction vs PaintNet is 93.2%, not 98.2%; vs Point-Wise there is no positive average because Containers is 2.0x worse (622.16 vs 313.98). Jerk: the average reduction vs PaintNet is 85.7%; the 97.8% figure comes only from the selective comparison with Point-Wise, and even that is not 97.0%. Section IV.A.b acknowledges only the Windows gain over Point-Wise and omits the coverage losses. Since the abstract's central claim rests on these numbers, the headline result is internally inconsistent with the only quantitative comparison.","section":"Abstract; Table I; Sec. IV.A.b"},{"comment":"The proposed model is conditioned on a 4-step, 24-DoF trajectory-history vector; the baselines are not. The ablation in Table III shows that this channel carries nearly the entire PCD gain on Windows: Previous Traj. gives 10.41, No Traj. gives 246.07, the latter being worse than the Point-Wise baseline's 55.71 (Table I). The paper does not state whether the history used at evaluation is the ground-truth prefix or model-generated. If it is the ground-truth prefix, the comparison is teacher-forced and not equivalent to the baselines; if it is model-generated, the protocol must still be applied to baselines. Without baselines that receive an equivalent conditioning signal, the margins in Table I cannot be attributed to the diffusion model or geometry encoder.","section":"Sec. III.C.a, Eqs. (3)-(4); Table III"},{"comment":"A stated contribution is avoiding 'post-hoc heuristic ordering or stitching' via 'simple sequential concatenation.' However, Sec. III.B says 'Segments are aligned by matching end and start poses, and concatenated into the final trajectory.' Matching endpoints is a post-hoc alignment heuristic, similar to PaintNet's overlapping-endpoint stitching. The manuscript should clarify what matching involves and whether this heuristic is used in the reported numbers; as written, the central distinction from the baseline is contradicted.","section":"Sec. III.B, Eq. (8); Introduction"},{"comment":"The coverage metric depends on the hand-set spray radius r_spray=0.05 m. No sensitivity analysis is reported, and with a single radius the metric may not discriminate among methods or could reverse the ranking at other plausible radii. Additionally, although the outputs are claimed to be 6-DoF, PCD and coverage use only 3D positions; orientation is never evaluated. A spray-coverage metric for 6-DoF trajectories should at least report whether the orientation is consistent with the surface normal, otherwise the '6-DoF' claim is unverified.","section":"Sec. IV.b, Eqs. (10)-(13)"},{"comment":"The paper claims a single policy generalizes across categories without category-specific training, but the experimental section does not state whether Table I comes from one joint model or four per-category models. If separate models were trained for each category, the generalization claim is not tested. In addition, Sec. III.C.b mentions a 'modified data processing pipeline' with category-specific scaling; it is not stated whether the baseline numbers were recomputed under this pipeline or taken from the original publications. Both points need to be disclosed before the comparison can be interpreted.","section":"Sec. III.C.b; Sec. IV.A"}],"minor_comments":[{"comment":"The two abstracts report different coverage numbers: '+67.5 percentage points' and 'increasing surface coverage by 61%'. Please make the claim consistent.","section":"Abstract"},{"comment":"The opening calls PCD 'our primary evaluation metric,' while Sec. IV.b calls coverage 'a primary evaluation metric.' Clarify which is primary.","section":"Sec. IV.A"},{"comment":"Smoothness is described as 'jerk statistics' but no definition, equation, or units are given.","section":"Sec. IV.b"},{"comment":"The text mentions a 'Random Traj.' variant with PCD 264.89, but Table III lists only Zero Traj., No Traj., and Previous Traj. Add the missing row or remove the mention.","section":"Sec. III.C.a; Table III"},{"comment":"The caption says 'Columns (left to right) show PaintNet, and 3DCovDiffusion (Ours),' but the figure header includes Multi-path prediction as a middle column. The caption is inconsistent with the figure.","section":"Fig. 4"},{"comment":"The code link is a placeholder ('can be found here') with no URL. Please provide the actual link or state that code will be released.","section":"Sec. I; end of paper"},{"comment":"The text quotes PointNet area-weighted coverage as 98.85%, while Table II gives 98.83%. Align the numbers.","section":"Table II"}],"recommendation":"reject","confidential_remarks":"For the editor: the paper is not ready for publication. The abstract's headline numbers are contradicted by Table I, and the main comparison is confounded by an unmatched trajectory-history conditioning channel. I recommend rejection unless the authors can correct the arithmetic, retrain fair baselines, and rerun the evaluation with a clearly specified protocol."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe one thing to know: this paper is not ready for prime time. The method itself — a DDIM-based diffusion policy conditioned on a point cloud plus a short trajectory history to generate 6DoF spray paths — is a legitimate engineering idea, and the ablations on the point-cloud encoder are informative. But the abstract's headline numbers are contradicted by the paper's own Table I, and the main comparison is confounded by an input channel the baselines never get.\n\nWhat's actually new: applying 3D Diffusion Policy-style conditioning to coverage path planning, with an ordered-sequence formulation that avoids the segment-wise prediction of PaintNet. The encoder ablation (Table II) is clean. The trajectory-conditioning ablation (Table III) is also valuable, even though it cuts against the authors: removing the trajectory history changes Windows PCD from 10.41 to 246.07, a ~24x swing. That tells you where the performance actually comes from.\n\nThe soft spots are not minor. First, arithmetic: the abstract promises 98.2% average PCD reduction, 97.0% jerk reduction, and +67.5pp coverage. From Table I, vs PaintNet the PCD reduction averages 93.2%, jerk reduction 85.7%, coverage gain 58.2pp. Vs Point-Wise, coverage is on average 11.4pp worse, and the model loses on three of four categories. The text in IV.A.b only mentions the Windows gain. That's an internal inconsistency, not a stylistic quibble. Second, the trajectory-history conditioning is an unmatched signal: no baseline gets the last 4 steps of ground truth. Table III shows it's load-bearing. Third, the paper claims to avoid stitching, but Eq. 8 still does endpoint alignment and concatenation. Fourth, the Containers results are poor (PCD 622 vs Point-Wise's 314), and the authors concede this in the limitation section while the introduction still claims data-scarce generalization. Also, the smoothness metric is never defined beyond 'jerk statistics', and only 3D positions are scored despite 6DoF claims.\n\nNone of this makes the method worthless. The core idea is sensible and the code release would help. But the paper as written doesn't support its claims.\n\nMy recommendation: send to peer review, but expect major revision. Referees should demand corrected numbers, matched baselines, a defined smoothness metric, and orientation error analysis. If the authors fix those, there's a usable paper here. As is, I would not cite it.","headline":"The architecture is a reasonable engineering contribution, but the abstract's numbers don't survive contact with the paper's own Table I, and the unmatched trajectory-history conditioning makes the main comparison unfair.","tokens_in":13438,"tokens_out":3164,"would_cite":false,"duration_ms":26465,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A single diffusion policy can generate ordered, smooth, high-coverage 6-DoF spray-painting trajectories from raw point clouds across different object categories, the paper argues.","keywords":["coverage path planning","diffusion policy","6-DoF trajectory generation","point cloud conditioning","spray painting","imitation learning","trajectory diffusion","robotic surface processing"],"falsifier":"Retrain PaintNet or the point-wise baseline with the same four-step ground-truth history as an additional input, or retrain the proposed model without the history channel, on the same train/test split; if the Chamfer-distance gap shrinks from roughly 98% to near parity, the central claim of a superior diffusion-based coverage policy is falsified. Additionally, recompute coverage with spray radius r=0.02 m and r=0.10 m to test whether the coverage gain persists.","tokens_in":12251,"feed_emoji":"🎨","tokens_out":3539,"duration_ms":31409,"temperature":0.7,"pith_summary":"The paper tries to establish that coverage path planning for industrial surface processing can be reformulated as conditional sequence generation and solved with one geometry-conditioned diffusion policy. Instead of predicting unordered stroke segments and stitching them with heuristics, the model denoises a full trajectory chunk conditioned on a point cloud and recent motion, producing ordered chunks that concatenate directly. If correct, this removes category-specific training and heuristic post-processing, letting a single policy serve cuboids, windows, shelves, and containers. The reported gains are large—about 98% lower point-wise Chamfer distance and jerk, and much higher surface coverage than prior learning-based baselines. The paper's own ablation, however, shows the model is trained with a ground-truth trajectory-history input that no baseline receives, which complicates the fairness of the headline comparison.","feed_headline":"Diffusion policy draws ordered spray paths from point clouds","feed_subtitle":"A single policy generalizes across four object classes; unmatched history input explains much of the gap.","key_machinery":"A conditional denoising diffusion implicit model (DDIM) with a cosine noise schedule, conditioned by a fused 128-dimensional vector. The condition concatenates (a) a global point-cloud descriptor from a three-layer MLP with max pooling (the 3DCovDiff encoder) and (b) a 64-dimensional embedding of the last four 6-DoF poses from a two-layer MLP. The condition is injected into every denoising residual block via FiLM (feature-wise linear modulation), and the training loss masks noise prediction over variable-length trajectories. This mechanism carries the whole claim: iterative denoising produces smooth, ordered, variable-length chunks, while the geometry encoder is what lets one policy generali","core_discovery":"The central claim is that coverage path planning can be cast as conditional sequence generation and solved by a geometry-conditioned denoising diffusion implicit model (DDIM). The policy consumes a partial point cloud of the target surface plus the most recent four end-effector poses and outputs the next trajectory chunk in temporally ordered form; chunks are concatenated sequentially without heuristic sorting or overlapping-endpoint alignment. Trained once on demonstrations for cuboids, windows, shelves, and containers, the same policy generalizes across those categories, and the paper reports large improvements over PaintNet, multi-path prediction, and point-wise baselines on Chamfer dista","pith_inferences":["The 98% Chamfer-distance and jerk margins over PaintNet likely overstate the diffusion contribution: the proposed model alone receives the last four ground-truth poses as input at train time, while no baseline gets any trajectory-history channel; a fair comparison would give the same conditioning to the baselines or remove it from the proposed model.","Only the 3D positions are scored by PCD and coverage; orientation is part of the 6-DoF output, so orientation accuracy is unverified. A test that separates position and orientation error could reorder the comparisons.","The coverage metric counts a face covered only if a segment passes within the hand-set spray radius (0.05 m) of its centroid; sweeping this radius would show whether the coverage advantage is robust or an artifact of the threshold.","A natural testable extension: condition the same diffusion policy on a short history of the model's own predictions instead of ground truth at training time, or evaluate closed-loop generation, to see whether the ordering and smoothness benefits survive without oracle context."],"forward_implications":["Industrial coating, painting, and polishing could use one learned policy per task instead of per object geometry, reducing manual redesign when the part changes.","Temporally ordered trajectory chunks can be concatenated sequentially, removing heuristic stitching and overlapping-endpoint alignment from the pipeline.","Because the policy outputs a distribution over trajectories, it can represent multiple valid coverage patterns for the same geometry, matching the multimodality of expert demonstrations.","The conditioning on recent poses makes the policy auto-regressive, so longer trajectories can be generated by feeding predictions back as context.","The 88-sample container category shows data scarcity, not model capacity, is the main brake on generalization to complex geometries."],"fun_headline_variants":["Diffusion draws ordered coverage paths from point clouds","Single diffusion policy handles multiple geometries for coverage","Geometry-conditioned diffusion plans coverage without stitching","Coverage path planning gets 98% lower Chamfer error via diffusion"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The load-bearing premise is that the comparison with baselines is fair, yet the proposed model is trained with a ground-truth trajectory-history input that no baseline receives, and an unevaluated re-processing of the dataset plus a hand-set spray radius in the coverage metric further affect the reported margins.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion draws ordered coverage paths from point clouds","Single diffusion policy handles multiple geometries for coverage","Geometry-conditioned diffusion plans coverage without stitching","Coverage path planning gets 98% lower Chamfer error via diffusion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000599,"raw_usage":{"total_tokens":2603,"prompt_tokens":673,"completion_tokens":1930,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":417,"completion_tokens_details":{"reasoning_tokens":1868}},"tokens_in":417,"tokens_out":1930,"duration_ms":13637,"temperature":1.0,"reasoning_tokens":1868,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T12:39:48.402599+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain PaintNet or the point-wise baseline with the same four-step ground-truth history as an additional input, or retrain the proposed model without the history channel, on the same train/test split; if the Chamfer-distance gap shrinks from roughly 98% to near parity, the central claim of a superior diffusion-based coverage policy is falsified. Additionally, recompute coverage with spray radius r=0.02 m and r=0.10 m to test whether the coverage gain persists.","supporting_citations":[],"review_version":1}