{"id":"fe0dae30-0ffb-4dc7-9f78-9a79865ebfe2","arxiv_id":"2506.13425","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A camera and inertial sensor filtering pipeline selects the most visible, highest bricks in a stack and estimates their 6DoF poses, outperforming simple baselines on a new synthetic benchmark.","lead":"This paper presents a system that helps a robot choose which brick to grab from a stack or pallet, and estimates the brick's exact position and orientation. The method is designed for construction and warehouse settings, where objects are stacked in orderly piles rather than loose bins.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Ground-truth graspability uses the same top-layer, high-visibility logic as the selection filter, so the benchmark may partly reward the method for matching its own labeling rule.","rationale":"The reader identifies the ground-truth proxy as the weakest assumption, and I concur: Algorithm 1's free-direction count is a visibility/topology proxy that closely tracks the method's own visibility-plus-height selection logic, so the reported mAP may partly measure self-consistency. This is the most load-bearing concern because the paper's stated contribution is selecting physically graspable objects from stacks, yet the only quantitative evidence is against a geometric proxy. The paper is otherwise internally coherent: the problem formulation, notation, and pipeline are clear, and the ablation comparisons to random, detector-confidence, inertial-only, and vision-only baselines are meaningful. I would not recommend rejection, since the paper is transparent about the synthetic evaluation and presents a reasonable first benchmark; however, the claim should be conditional on a physical validation or an explicit reframing of the ground truth as a visibility/topology criterion rather than graspability.","tokens_in":8031,"tokens_out":1119,"duration_ms":9984,"concrete_test":"Run real-robot grasp trials on a physical brick stack with the proposed pipeline, recording grasp success versus failure for top-k=1, and compare against the synthetic ground-truth labels for the same scenes; if agreement between predicted graspable candidates and physical success is significantly lower than the synthetic mAP, the benchmark overstates graspability.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that F = F_v composed with F_h selects graspable objects accurately (mAP 0.79 at top-k=1). The benchmark's ground-truth graspability (Algorithm 1) classifies a brick as graspable if it has at least four free principal directions, or exactly three with no upper neighbor and at least one free lateral direction in both X and Y. This rule is essentially a geometric proxy for being on or near the top layer with lateral clearance. The proposed filter uses the same signal: visibility ratio r_i (Eq. 3) with threshold 0.80 plus IMU height ranking. A brick that satisfies the ground-truth rule is exactly a brick that tends to be visible and high in the stack, so the evaluation measures agreement with the labeling rule, not physical grasp success. The paper provides no real-robot success rates or physical grasp trials with the proposed metric; the deployment section is qualitative and in a supplementary video. The synthetic benchmark validates the filter against its own assumption, leaving the physically load-bearing claim unverified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper formally introduces the task of selecting an object to grasp from a structured stack and estimating its 6DoF pose. The proposed pipeline runs YOLOv8 detection, ZebraPose 6DoF estimation, a vision-based filter using the modal/amodal mask visibility ratio of Eq. (3) with threshold epsilon_vis=0.80, and an IMU-based gravity-axis height filter that keeps the top-k candidates. The authors contribute a BlenderProc synthetic brick-stack dataset with graspability labels generated by a free-neighbor-direction rule (Algorithm 1), an evaluation protocol that averages precision over ADD-S and MSSD pose thresholds, and ablations showing mAP 0.79 for top-k=1 versus 0.73 for vision-only and 0.61 for inertial-only filtering. A qualitative real-robot deployment on a brick pallet is shown in the supplementary video.","tokens_in":8205,"tokens_out":6780,"duration_ms":73636,"significance":"If the claims hold, the task definition and the dataset fill a genuinely under-explored gap in robotic manipulation of structured stacks, and the idea of combining a visibility mask with IMU-based height ranking is simple and practical. The evaluation metric that couples object selection with pose accuracy is a useful contribution to benchmarking this class of problems. However, the main quantitative claim currently rests on a synthetic benchmark whose ground-truth graspability labels are generated by a hand-coded geometric rule that closely resembles the method's own filters, and the real-world deployment is only shown qualitatively. These issues, together with the absence of error bars and sensitivity analysis, mean that the central claim is plausible but not yet convincingly established.","major_comments":[{"comment":"The ground-truth graspability label is generated by the free-neighbor rule in Algorithm 1, which labels a brick graspable if it has at least four missing principal directions (or, per the text, three with no upper neighbor and lateral clearance). This is essentially the same top-layer, unobstructed-object notion that the proposed visibility threshold (Eq. 3) and IMU height top-k filter (Section IV.E) implement. The mAP values in Tables I and II therefore largely measure how well the filter reproduces the labeling rule, not how often a grasp would physically succeed. To support the central claim, the authors should add an independent validation, for example real-robot grasp success rates on stacks, or relabel the benchmark with a physically motivated criterion such as collision-free force-closure grasps and show that the method ranking is unchanged.","section":"V.B / Algorithm 1 vs. IV.D-IV.F"},{"comment":"As printed, Algorithm 1 is not a faithful implementation of Section V.B. Line 15 increments m when every queried neighbor has positive IoU with the shifted bounding box, which counts a direction as blocked, whereas the text says a direction is missing precisely when no neighbor overlaps. The pseudocode also omits the three-missing-directions case with the additional structural constraints described in Section V.B. If this is a typesetting error, the benchmark code must be released and the labels regenerated; otherwise the ground-truth labels may be inverted.","section":"Algorithm 1, lines 15-18"},{"comment":"Equation (7), o_i,z_g = alpha_s o_s, is dimensionally inconsistent because alpha_s is an acceleration vector and o_s is a position vector in the sensor frame. The height projection should be the scalar product of the object position with the normalized gravity direction, for example (alpha_s / ||alpha_s||)^T o_s. The missing normalization and missing dot-product notation need to be corrected, and the implementation of the height ranking should be clarified, since this equation is load-bearing for the IMU-based filter.","section":"IV.E, Eq. (7)"},{"comment":"The quantitative evaluation reports single scalar mAP values without error bars, confidence intervals, or multiple seeds, despite the use of learned components and a 1K-image test split. The differences that motivate the combined filter are modest (mAP 0.79 vs. 0.73 vs. 0.61 in Table II), and without variance estimates it is unclear whether the combined method is significantly better than vision-only filtering. In addition, epsilon_vis=0.80 is fixed without a sensitivity analysis; a sweep over epsilon_vis and over top-k is needed to show the result does not depend critically on the chosen threshold.","section":"Tables I and II / VI.C-VI.D"}],"minor_comments":[{"comment":"The deployment section refers only to a supplementary video. Please provide quantitative real-world measurements, such as the number of pick attempts and successes or pose errors against manual annotation, so that the real-world claim can be assessed.","section":"VI.E"},{"comment":"The dataset is described as being made available upon acceptance, but no URL or benchmark download mechanism is given. Please include a concrete release mechanism so that the benchmark is actually usable by other researchers.","section":"V.A / VI.B"},{"comment":"The composition notation F = F_v composed with F_h is used, but the accompanying sentence says the inertial filter is applied first and the vision filter second. Please state the order explicitly in both the notation and the text so there is no ambiguity about which filter is applied first.","section":"IV.F"}],"recommendation":"major_revision","confidential_remarks":"The main risk is that the synthetic benchmark's graspability labels are defined by the same top-layer-visibility heuristic as the proposed filters, so the headline mAP may partly measure self-consistency. The paper would be substantially strengthened by real-robot grasp trials with success rates, error bars over seeds, and a sensitivity analysis for epsilon_vis. If the authors provide those, I would be willing to support acceptance; in the current form the central empirical claim is not yet fully supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid, well-scoped paper about a real subproblem (grasping from structured stacks), with a clean benchmark and a simple method that beats its baselines. The catch is that the benchmark's ground-truth graspability is a geometric proxy that partly uses the same signal as the method, so the performance gap is partly self-confirmation. That doesn't kill the paper, but it means the headline numbers shouldn't be read as physical grasp success.\n\nWhat's actually new: the formal problem definition (selecting and posing objects from stacks), the procedurally generated brick-stack dataset, and the evaluation metric that combines selection quality with pose accuracy. The method itself—visibility-ratio thresholding plus IMU height ranking—is a straightforward combination of standard pieces, but it's applied to a less-studied setting and the ablations are clear. Tables I and II show a big jump from 0.17 mAP for unfiltered poses to 0.79 for the combined filter at top-k=1, with vision-only at 0.73 and inertial-only at 0.61. The paper is honest in the abstract and conclusion that the problem remains hard for error-free performance.\n\nThe soft spots are real. Algorithm 1 labels a brick graspable if it has at least four free neighbor directions (or three with no upper neighbor and free laterals). That's essentially a 3D way of saying \"on the top layer with surrounding clearance.\" The method's visibility ratio and height ranking are 2D/1D proxies for the same idea. So the benchmark and the method are not independent. This is a common issue in synthetic benchmarks, but here it's more direct because the labels are hand-coded rather than learned or physically tested. There are no real-robot success rates, no error bars or multiple seeds, and epsilon_vis=0.80 is fixed without sensitivity analysis. The dataset and code are promised \"upon acceptance\" but not yet available. Also, Eq. (7) is garbled—the height extraction from the rotation is written as acceleration times something; it's a typo, but it makes that part hard to follow.\n\nNone of this is fatal. The contribution is the problem setup plus a usable benchmark, and the method is a sensible first baseline. But the evaluation currently validates the filter against its own assumption, not against physical grasp success. If I were handling this, I'd send it to reviewers and ask for a real-robot quantitative result or at least a discussion of the label proxy's limitations, plus release the data.\n\nWho it's for: people working on robotic manipulation from structured scenes, especially construction or warehouse palletizing. It's worth a careful read for that community, and the dataset could be a useful testbed if released. My recommendation: accept with revisions, not desk reject.","headline":"A solid benchmark for stack grasping, but the ground truth labels partly mirror the method's own selection logic, so treat the headline numbers as self-consistent rather than physical.","tokens_in":8739,"tokens_out":2835,"would_cite":false,"duration_ms":28540,"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":"A visual-inertial filter that keeps top-layer, unobstructed bricks reaches 0.79 mAP for single-brick grasping from a stack, versus 0.17 with no filtering","keywords":["robotic grasping","6DoF pose estimation","object selection from stacks","visual-inertial filtering","synthetic benchmark dataset","brick picking","visible mask ratio","top-k candidate selection"],"falsifier":"Recompute the pipeline's mAP against graspability labels produced by a physical grasp simulator for the two-prong gripper, checking collisions, friction, and post-grasp stability instead of free-direction space; if mAP falls close to the vision-only or unfiltered values, the reported gain came from matching the labeler's own criterion rather than from true grasp readiness.","tokens_in":7802,"feed_emoji":"🧱","tokens_out":9885,"duration_ms":88146,"temperature":0.7,"pith_summary":"The paper argues that grasping from a structured stack requires a second decision beyond pose estimation: choosing which object to grab. It proposes a camera-plus-IMU filter that keeps only bricks that are both highly visible (the visible part of the object nearly covers its full silhouette) and near the top of the stack (highest positions along gravity), then estimates the six-degree-of-freedom pose of the chosen brick with an RGB pose estimator. To benchmark this, it introduces a procedurally generated synthetic dataset of stacked bricks with graspability labels and an average-precision metric that counts a selected pose as correct only if it matches a labeled graspable pose within ADD-S and MSSD thresholds. On that benchmark, the combined filter reaches 0.79 mean average precision for a single pick, against 0.17 with no filtering, 0.60 with detector-confidence selection, and 0.61 with inertial-only filtering. A deployment on a construction-site robot arm demonstrates the pipeline in practice.","feed_headline":"A vision+IMU filter picks the right brick with 0.79 mAP","feed_subtitle":"For single grabs from a brick stack, visibility plus gravity-height beats detector confidence and inertial-only baselines.","key_machinery":"The load-bearing mechanism is the visibility ratio $r_i = |M^i_m| / |M^i_a|$, computed by rendering the full amodal silhouette of the estimated pose and dividing its pixel area by the visible segmentation mask, combined with a gravity-axis height rank obtained by projecting each object center into the global frame using the IMU acceleration reading. The two filters are composed as $F = F_v \\circ F_h$, with visibility threshold $\\epsilon_{vis}=0.80$ and top-$k$ height selection. On the benchmark side, a KD-tree and axis-aligned bounding-box IoU check labels a brick graspable when at least four of six principal directions are free (or three with no upper neighbor plus one free lateral and one free front-back direction), and evaluation uses Hungarian matching and average precision over ADD-S and MSSD threshold sets. These two quantities carry the argument.","core_discovery":"The central claim is that object selection for stack grasping should be treated as a filtering problem on raw six-degree-of-freedom pose hypotheses rather than as a ranking by detector confidence. The paper's specific proposal is the composed filter $F = F_v \\circ F_h$: first rank all detected bricks by their height along the IMU-measured gravity axis and keep the top $k$, then discard any brick whose visibility ratio $r_i = |M^i_m| / |M^i_a|$ falls below the threshold $\\epsilon_{vis}=0.80$. On the synthetic benchmark with $k=1$, this yields 0.79 mAP over ADD-S and MSSD thresholds, versus 0.17 for all unfiltered poses, 0.60 for detector-confidence selection, 0.61 for inertial-only filtering, and 0.73 for vision-only filtering. The paper also claims that the problem remains genuinely hard: selecting two or three bricks from one image lowers mAP to 0.70 and 0.65, and a completely error-free solution is not achieved.","pith_inferences":["An independent grasp-planner label set, using collision meshes, gripper geometry, and friction, would test whether the free-direction rule is a good proxy for graspability; until then, the mAP numbers should be read partly as self-consistency between the selector and the labeler.","The visibility-ratio idea transfers to any object category with amodal predictions, and a network that regresses $r_i$ directly from RGB could remove the need to render amodal masks at inference time.","Because the IMU contributes only the gravity direction for height ranking, an image-only vertical-direction estimate from vanishing points or floor segmentation could make the method work without an IMU, at some cost to accuracy."],"forward_implications":["A robot can select a single graspable brick from a stack with an RGB camera, an IMU, a detector, a pose estimator, and the composed filter, and the construction-site deployment is the paper's evidence for this.","The filter is generic over the detector and pose estimator, so replacing either module with a more accurate one should improve the pipeline without changing the selection logic.","The mAP drop from 0.79 for top $k=1$ to 0.70 for $k=2$ and 0.65 for $k=3$ implies the method is best suited to single-object picking, and multi-object picking from stacks remains open.","Visibility filtering alone accounts for most of the gain (0.73 mAP) while inertial filtering alone gives 0.61, but the two are complementary and together add roughly six points of mAP."],"supporting_citations":[{"why":"Supplies the RGB-only six-degree-of-freedom pose hypotheses and visible segmentation masks that the filtering pipeline consumes.","marker":"[8]"},{"why":"Defines the MSSD error metric, threshold sets, and evaluation conventions used for pose-correctness judgement.","marker":"[9]"},{"why":"Provides the 2D bounding-box detections that crop regions of interest for the pose estimator.","marker":"[14]"},{"why":"Procedurally generates the synthetic stacked-brick dataset with RGB, depth, masks, and six-degree-of-freedom pose annotations.","marker":"[15]"},{"why":"Defines ADD-S, the symmetry-aware pose error metric used to decide whether a selected pose matches ground truth.","marker":"[16]"},{"why":"Provides the Hungarian assignment algorithm that pairs predicted candidate poses with ground-truth graspable poses for precision computation.","marker":"[17]"}],"fun_headline_variants":["Height and visibility filter beats confidence for brick picking","Stack grasping: filter pose hypotheses by height and visibility","0.79 mAP brick selection by filtering pose hypotheses","Filtering pose hypotheses improves brick pick accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The benchmark defines 'graspable' by a geometric rule, free space around a brick in at least four directions, and the method selects exactly the highly visible, top-layer bricks that this rule favors, so the evaluation may measure how well the filter matches its own definition rather than how often a real gripper succeeds.","fun_headline_variants_meta":{"raw":{"variants":["Height and visibility filter beats confidence for brick picking","Stack grasping: filter pose hypotheses by height and visibility","0.79 mAP brick selection by filtering pose hypotheses","Filtering pose hypotheses improves brick pick accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000784,"raw_usage":{"total_tokens":3449,"prompt_tokens":920,"completion_tokens":2529,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":536,"completion_tokens_details":{"reasoning_tokens":2468}},"tokens_in":536,"tokens_out":2529,"duration_ms":18183,"temperature":1.0,"reasoning_tokens":2468,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:02:00.636152+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the pipeline's mAP against graspability labels produced by a physical grasp simulator for the two-prong gripper, checking collisions, friction, and post-grasp stability instead of free-direction space; if mAP falls close to the vision-only or unfiltered values, the reported gain came from matching the labeler's own criterion rather than from true grasp readiness.","supporting_citations":[{"cited_title":"Zebrapose: Coarse to fine surface encoding for 6dof object pose estimation,","cited_arxiv_id":null,"evidence_quote":"Supplies the RGB-only six-degree-of-freedom pose hypotheses and visible segmentation masks that the filtering pipeline consumes."},{"cited_title":"Bop challenge 2023 on detection segmentation and pose estimation of seen and unseen rigid objects,","cited_arxiv_id":null,"evidence_quote":"Defines the MSSD error metric, threshold sets, and evaluation conventions used for pose-correctness judgement."},{"cited_title":"Blenderproc2: A procedural pipeline for photorealistic rendering,","cited_arxiv_id":null,"evidence_quote":"Procedurally generates the synthetic stacked-brick dataset with RGB, depth, masks, and six-degree-of-freedom pose annotations."},{"cited_title":"Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes,","cited_arxiv_id":null,"evidence_quote":"Defines ADD-S, the symmetry-aware pose error metric used to decide whether a selected pose matches ground truth."}],"review_version":2}