{"id":"4aebb39e-6170-4115-bfdb-ae970675f5fb","arxiv_id":"2506.14066","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A strawberry harvesting pipeline that completes partial point clouds and treats other berries as obstacles reports 79.17% grasp success and cuts obstacle hits from 43.33% to 13.95% in a lab.","lead":"This paper describes a robotic strawberry-picking pipeline that reconstructs the hidden parts of berries from a single camera view before deciding where to grasp. In lab trials with five berries, the system reports a higher grasp success rate and fewer collisions with nearby fruit than the comparison method.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported grasp improvements do not isolate point cloud completion as the causal driver, because the only external baseline uses a different robotic platform and the internal comparison does not ablate completion separately.","rationale":"The reader's weakest assumption focuses on the fixed-shape/size CAD assumption and generalization to variable field strawberries. That is a legitimate limitation, and the authors themselves acknowledge it in Section III ('prior fixed shape and size') and Section VII (future work on 'adaptive shape completion methods that can handle greater variability'). My concern is more immediate and internal to the reported evaluation: even within the controlled lab setting, the experimental design does not demonstrate that point cloud completion is responsible for the grasp improvements. The only external baseline is from a different robot and environment, so the comparison is confounded. The internal Table III comparison is not clearly specified as a completion ablation, and the success-rate numbers are presented without confidence intervals or raw counts. These are not disputes with external consensus; they are checks on internal validity. A 2×2 ablation with proper statistics would directly settle whether the completion module earns its central role. If the ablation shows no completion effect, the paper's core contribution collapses and a REJECT would be warranted; if it confirms a completion effect, the conditional acceptance is appropriate. Since the current evidence leaves that question open, the CONDITIONAL verdict remains the right one.","tokens_in":9057,"tokens_out":5696,"duration_ms":66354,"concrete_test":"On the same robot, same lab scene, and fixed strawberry arrangement, run a 2×2 ablation crossing completion on/off with obstacle-aware planning on/off, with at least 80 trials per cell and bootstrapped 95% confidence intervals for rho_s and rho_h. If completion-on/planning-on is statistically indistinguishable from completion-off/planning-on, then point cloud completion is not load-bearing for the reported success; the paper should then be revised to claim obstacle-aware motion planning as the main contribution.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that integrating point cloud completion with obstacle-aware planning yields the reported rho_s = 79.17% and rho_h reduction from 43.33% to 13.95%. For that claim to hold, the completion module must be a causal contributor to the improvement, and the comparisons must rule out platform and environment confounds. Neither condition is met. Table I compares the full pipeline with [9], which uses a different robot, gripper, camera, and a vertical-farming environment (Section V.E), so the 26.67% relative improvement cannot be attributed to the proposed method alone. Table III is the only internal comparison, but the text describes toggling 'our method' and 'obstacle hit rate' without specifying whether completion is disabled in the × condition; if × only removes the obstacle-aware planner, then point cloud completion is never ablated in the grasping evaluation. Table II supports only reconstruction accuracy, not grasping success. Finally, no confidence intervals or raw counts are reported for the 80 trials, so the headline differences may be within sampling noise.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an end-to-end robotic strawberry harvesting pipeline that combines point cloud denoising/segmentation, PointAttn-based point cloud completion, target selection based on ripeness and distance, and occupancy-map-aware motion planning that treats non-target strawberries as obstacles. The authors report a Chamfer Distance of 1.10 mm for shape completion, a grasp success rate of 79.17% with a success-to-attempt ratio of 89.58%, and a reduction in obstacle hit rate from 43.33% to 13.95% over a baseline. The evaluation is based on 80 real-world lab trials with five strawberries and a Unitree Z1 arm, with code publicly available.","tokens_in":9270,"tokens_out":3893,"duration_ms":43746,"significance":"If the claims hold, the paper makes a useful contribution to agricultural robotics by demonstrating that combining point cloud completion with obstacle-aware planning can improve grasping of partially occluded fruit. The strengths are the public code release, a real-robot evaluation, and a clean pipeline description that others can reproduce. However, the significance is currently conditional: the grasping comparison uses a different robotic platform and environment, the ablation in Table III does not isolate the completion module, the Chamfer Distance evaluation lacks a clear common protocol, and no statistical backing is provided for the claimed improvements. The stated assumption of a fixed strawberry shape and size further limits the generality of the results, although the authors acknowledge this in the future-work section.","major_comments":[{"comment":"The only grasping baseline, Shape Completion [9], was evaluated on a different robot, gripper, camera, and vertical-farm environment, which the authors acknowledge. Consequently, the reported 26.67% relative improvement in ρs is not attributable to the proposed method alone. A same-platform baseline or an ablation of the proposed planning/completion components on the authors' own robot is needed to support the central claim that the pipeline improves grasping.","section":"§V.E, Table I"},{"comment":"The 'Ours Used ×' row does not specify which components are disabled. If the × condition removes only the obstacle-aware planner while point cloud completion remains active, then the reduction in obstacle hit rate from 43.33% to 13.95% provides no evidence about the contribution of completion to grasping, and completion is never ablated in the grasping evaluation. The paper must define the ablated configuration precisely and ideally include separate ablations for completion and obstacle-aware planning. In addition, no raw trial counts, confidence intervals, or statistical tests are reported for the 80 trials, so the word 'significantly' in the abstract and Section VI is unsupported.","section":"§VI, Table III"},{"comment":"The Chamfer Distance comparison lacks a common evaluation protocol. It is not stated whether the baselines [24] and [32] were run on the same partial inputs, the same ground truth, and the same evaluation split as the proposed method. Since the lab-collected real data were manually annotated, combined with simulation data for training, and also used for the reported CD evaluation, the paper must disclose whether the strawberries used for the CD test overlap with the training set; otherwise the 1.10 mm result may reflect memorized shapes rather than genuine completion capability.","section":"§V.A, §VI, Table II"}],"minor_comments":[{"comment":"The voxel size vs and the minimum-points-per-voxel threshold of 30 are given without units and without justification. Also, the set-builder notation 'pj∈ p| floor(pj/vs)=i' is malformed; please write the condition over voxel indices explicitly.","section":"§IV.A, Eqs. (4)-(5)"},{"comment":"The detection and segmentation modules YOLOv8 and SAM2 are described only by name; no detection or segmentation accuracy metrics are reported, and the paper does not explain how mask-point association is performed when multiple masks overlap or when detection misses a strawberry. This makes it hard to assess how perception errors propagate to grasp success.","section":"§V.C"},{"comment":"The occupancy map Mocc is generated only from obstacle point clouds, while the target's completed point cloud is used for grasp estimation. Please clarify how the target geometry is kept out of the collision map and how the completed target shape influences the planned trajectory.","section":"Algorithm 1, lines 14-17"},{"comment":"The denominators for ρa, ρs, and ρh are not explicitly stated. Reporting the raw counts (number of detections, attempts, successes, and obstacle hits) would allow the rates to be checked and would partially mitigate the absence of confidence intervals.","section":"Tables I and III"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's scope and has a reproducible code release, so I see no novelty or fit concern. The main risk is evaluative rigor: the baseline mismatch, the ambiguous ablation, and the unclear CD protocol are all fixable in revision, but they currently prevent the central claim from being accepted as stated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick read on arXiv:2506.14066. It's a legitimate incremental advance: they combine YOLOv8 detection, SAM2 segmentation, a PointAttn-based completion network, and collision-aware planning that treats non-target berries as obstacles. The integration is new, and they report real-robot lab trials with code released. The reconstruction CD of 1.10 mm is plausibly good, though I'd want to know the evaluation protocol.\n\nWhat it does well: the system is coherent, the obstacle-aware planning idea is a sensible extension over prior work like [9], and the lab setup is reasonable even if it's only five strawberries. The 80 trials show a grasp success of about 79% and a drop in obstacle hit rate from 43% to 14%, which is encouraging.\n\nThe soft spots are real. The main comparison to [9] uses a different platform and vertical-farm environment, so the 26.67% relative improvement over that baseline can't be attributed to any specific component. Table III is the only internal comparison, and from the text it's not clear whether the \"×\" condition disables completion or only the obstacle-aware planner; if completion stays on in both, then the completion module's contribution to grasping is never ablated. No confidence intervals or raw counts are provided, so the headline differences could be within sampling noise for 80 trials. The fixed-shape strawberry assumption is also explicitly stated (Section III) and likely limits field transfer.\n\nNone of these are fatal. The paper does exactly what a systems paper should do: it builds something, releases code, and measures it on a physical robot. The citations to prior work look fair, and the claims in the abstract match the results as presented.\n\nWho should read this: people working on fruit-picking manipulation or sim-to-real pipelines for agriculture. It's not a methodological breakthrough, but it's a clean example of an integrated pipeline. I would not cite it in my own work unless I needed a recent agricultural grasping baseline.\n\nFor peer review: yes, send it out. But require the authors to clarify the Table III ablation, add variance or at least raw counts, and either run a true baseline on their own platform or soften the causal claims about completion.\n\nBest.","headline":"A competent integrated system for strawberry harvesting with real-robot numbers, but the evaluation doesn't prove the completion module is the cause of the improvement.","tokens_in":9755,"tokens_out":1759,"would_cite":false,"duration_ms":17273,"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":"This paper claims that completing partially occluded strawberry point clouds before grasp planning raises whole-fruit harvesting success rates and sharply reduces collisions with neighboring fruit.","keywords":["point cloud completion","robotic strawberry harvesting","occlusion handling","grasp planning","obstacle avoidance","occupancy map","Chamfer Distance","RGB-D perception"],"falsifier":"Run the same pipeline on strawberries whose sizes span a realistic range (for example, plus or minus 30% of the fixed CAD dimensions) in the same lab setup and measure the grasp success rate and Chamfer Distance. If the success-to-attempt ratio falls back toward the 74.07% baseline, or the completed point clouds deviate by more than about 2 mm once size deviates, the fixed-shape assumption is the weak link and the central claim loses its generality.","tokens_in":8880,"feed_emoji":"🍓","tokens_out":3804,"duration_ms":36733,"temperature":0.7,"pith_summary":"The paper argues that the missing piece in robotic strawberry harvesting is not detection alone but geometric reconstruction under occlusion. Its pipeline takes a single RGB-D view of a plant, isolates each strawberry, completes the partial point cloud of every fruit, and then plans a grasp toward the nearest ripe strawberry while treating all other strawberries as obstacles in an occupancy map. The claim, supported by 80 real-world trials, is that this combination yields a grasp success rate of 79.17%, a success-to-attempt ratio of 89.58%, a reconstruction error of 1.10 mm Chamfer Distance, and a drop in obstacle hit rate from 43.33% to 13.95%. A careful reader would care because it offers a concrete route to damage-free automated picking of a fragile crop in cluttered scenes.","feed_headline":"Completing occluded fruit shapes lifts harvest success to 79%","feed_subtitle":"Occluded strawberries get fully reconstructed in 3D, so the robot avoids bumping neighboring fruit while picking.","key_machinery":"The load-bearing object is the point cloud completion network, an attention-based model following PointAttn, which maps a denoised partial point cloud of a strawberry to progressively denser completed point clouds under a hierarchical Chamfer Distance loss. It is paired with an occupancy-map planner: the completed point clouds of non-target strawberries are unioned and discretized into a 3D occupancy grid, so inverse kinematics and trajectory planning treat neighboring fruit as obstacles while approaching the selected ripe target.","core_discovery":"The central claim is that a complete 3D reconstruction of every strawberry in a scene, not just the target, is what enables safe whole-fruit grasping under occlusion. For each detected strawberry the pipeline produces a dense completed point cloud from a noisy partial scan, selects the ripe strawberry closest to the end-effector, and builds an occupancy map from the completed point clouds of all other strawberries so the planned trajectory avoids them. In real-world trials this pipeline achieves a grasping success rate of 79.17% with a success-to-attempt ratio of 89.58%, a Chamfer Distance of 1.10 mm that is lower than the compared completion baselines, and a reduction in obstacle hit rate from 43.33% to 13.95% when the obstacle-aware planning is enabled.","pith_inferences":["If the central claim holds, then the same pattern of completing all objects in a scene before grasp planning should transfer to other roughly rigid fruits such as apples or tomatoes, provided the completion model is retrained on their shapes.","A natural ablation test would separate the two contributions: planning on partial point clouds with obstacle avoidance would isolate how much of the gain comes from completion itself, while planning on completed clouds without the occupancy map would isolate the obstacle-avoidance contribution.","The reported drop in obstacle hit rate suggests that completion improves obstacle localization as well as target reconstruction; quantifying those two effects separately would sharpen the claim.","Field deployment is the real test: the fixed-shape CAD assumption and lab lighting mean the reported numbers may not represent natural variability in berry size, ripeness deformation, and leaf clutter."],"forward_implications":["In cluttered scenes, completing the geometry of both target and non-target fruit reduces collisions; the reported obstacle hit rate drops by about 68% when the full pipeline is used.","The lower Chamfer Distance (1.10 mm vs 2.42 mm for the best baseline) translates into more reliable grasp localization, which is the stated reason for the higher grasping success rate.","Treating unripe and occluded strawberries as obstacles rather than discarding them allows the robot to harvest without damaging the surrounding crop.","Because the pipeline works from a single RGB-D view on an embedded GPU, it is compatible with field-style deployment conditions that allow only one sensor pass.","The success-to-attempt ratio of 89.58% compared to 74.07% for the shape-completion baseline indicates fewer failed attempts per successful harvest, which matters for cycle time in a real picking operation."],"supporting_citations":[{"why":"Supplies the attention-based point cloud completion architecture (PointAttn) that the paper adapts for strawberry shape completion.","marker":"[7]"},{"why":"Provides the closest baseline for grasping success and the prior 3D shape-completion harvesting pipeline that this work extends with obstacle awareness.","marker":"[9]"},{"why":"Serves as a Chamfer Distance baseline (CoRe) for strawberry shape completion in controlled environments.","marker":"[24]"},{"why":"Serves as the DeepSDF-based baseline for Chamfer Distance comparison in the completion experiments.","marker":"[32]"},{"why":"YOLOv8 is used to detect strawberries and provide bounding boxes that seed the segmentation stage.","marker":"[30]"},{"why":"SAM2 generates the precise instance masks used to isolate individual strawberry point clouds for completion.","marker":"[31]"}],"fun_headline_variants":["Occluded strawberries: 3D completion lifts grasp success to 79%","Filling in occluded fruit shapes lets robots grasp safely at 79%","Point cloud completion for occluded fruit boosts grasp rate to 79%","Robot strawberry picker: completing partial scans cuts obstacle hits by 68%","How a robot sees through leaf clutter to grasp strawberries at 79%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline assumes every strawberry has one fixed known shape and size from a CAD model, so if real fruit vary in size, shape, or ripeness deformation, the completed point clouds and the grasps built from them become inaccurate.","fun_headline_variants_meta":{"raw":{"variants":["Occluded strawberries: 3D completion lifts grasp success to 79%","Filling in occluded fruit shapes lets robots grasp safely at 79%","Point cloud completion for occluded fruit boosts grasp rate to 79%","Robot strawberry picker: completing partial scans cuts obstacle hits by 68%","How a robot sees through leaf clutter to grasp strawberries at 79%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00054,"raw_usage":{"total_tokens":2583,"prompt_tokens":933,"completion_tokens":1650,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":549,"completion_tokens_details":{"reasoning_tokens":1550}},"tokens_in":549,"tokens_out":1650,"duration_ms":13135,"temperature":1.0,"reasoning_tokens":1550,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:54:00.545502+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same pipeline on strawberries whose sizes span a realistic range (for example, plus or minus 30% of the fixed CAD dimensions) in the same lab setup and measure the grasp success rate and Chamfer Distance. If the success-to-attempt ratio falls back toward the 74.07% baseline, or the completed point clouds deviate by more than about 2 mm once size deviates, the fixed-shape assumption is the weak link and the central claim loses its generality.","supporting_citations":[{"cited_title":"Contrastive 3d shape completion and reconstruction for agricultural robots using rgb-d frames,","cited_arxiv_id":null,"evidence_quote":"Serves as a Chamfer Distance baseline (CoRe) for strawberry shape completion in controlled environments."},{"cited_title":"Panoptic mapping with fruit completion and pose estimation for horticultural robots,","cited_arxiv_id":null,"evidence_quote":"Serves as the DeepSDF-based baseline for Chamfer Distance comparison in the completion experiments."},{"cited_title":"Improving robotic fruit harvesting within cluttered environments through 3d shape completion,","cited_arxiv_id":null,"evidence_quote":"Provides the closest baseline for grasping success and the prior 3D shape-completion harvesting pipeline that this work extends with obstacle awareness."}],"review_version":1}