{"id":"8a93ecd4-d00b-4b30-bd70-061d3ab34551","arxiv_id":"2411.18133","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A training-free cross-device robotic grasping policy uses background-class softmax and geometry clustering to detect and grasp open-world objects.","lead":"This paper presents a plug-and-play policy that lets a robot arm grasp unfamiliar objects using a pre-trained 3D object detector, without retraining. The approach was tested on two different robot-camera pairs and is claimed to work across device brands and object categories.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SSGC-Seg's foreground/background rule assumes seen-class softmax exceeds background on unseen objects; no evidence supports this, so the open-category detection claim can collapse.","rationale":"Good-faith reading: the paper is a systems contribution—transfer a frozen closed-set 3D segmenter to new cameras and robots and grasp arbitrary objects by using background-softmax thresholding plus geometric clustering and a hand-crafted scoring rule. The mechanism is coherent and the real demos are credible. The load-bearing point is not the absence of PyTorch baselines or trial counts (important for evidence quality, but fixable with more data) but the binary-separation rule in Eq. (1)–(2). Everything downstream—clustering, ScoreNet‡, robot grasping—assumes the foreground mask is correct. If the closed-set softmax does not satisfy max(S[2:M]) > S[1] on the object points of a novel category, the object never enters the pipeline. The paper's only quantitative validation (Tab. 1, Boxes) uses the same seen classes the model was trained on, so it cannot probe this rule in the open world. The real-world tables report 100% recognition but on a small, undocumented number of objects and without per-point segmentation accuracy; a single success with a spray can does not characterize the failure boundary. Therefore we agree with the reader's weakest assumption. The concrete test—running the frozen model on a set of diverse unseen objects and measuring foreground recall—directly settles whether the assumption holds; if the objection is real, it will show up as missed objects. We thus keep the CONDITIONAL verdict: release code and weights, add the open-world segmentation stress test, and report trial counts, variances, and baselines.","tokens_in":12573,"tokens_out":7229,"duration_ms":66404,"concrete_test":"Take the released pretrained PBNet/Boxes model and the RealSense D455 setup, capture point clouds of at least 20 unseen tabletop objects (varied shape and texture, e.g., YCB objects plus plush toys, bottles, cans), and manually annotate object masks. Run SSGC-Seg (Eq. 1–2) and compute per-object foreground point recall and background precision. A pass requires all objects to have >80% foreground recall and no object entirely missed; also check that table points are not labeled foreground. If any object fails (e.g., the plush toy yields max(S[2:M]) < S[1] on most points), the assumption behind Eq. (2) breaks and the open-category claim is falsified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—plug-and-play open-category grasping across devices—hinges on SSGC-Seg correctly labeling every point as foreground or background via Eq. (1)–(2). The rule is: a point is foreground iff max(S[2:M]) > S[1], where S[1] is the background-class softmax. This is a hard threshold on a closed-set classifier's softmax behavior. The paper states the assumption as an observation but never characterizes when it holds. On the closed-set Boxes benchmark (Tab. 1) all objects are seen boxes, so the seen-class logits are likely confident and the rule works. In an open-world scene, an unseen object (toy, spray can, snack) does not activate any trained class; its maximum seen-class probability may be below the background probability, causing the entire object to be assigned to background, yielding no proposal and no grasp. Conversely, domain-shifted table/floor points may activate a seen class and be wrongly treated as foreground. The real-world demos (Tab. 2) show only a handful of objects with no per-point segmentation metrics, trial counts, or failure analysis; they cannot establish the rule across object geometries. Thus the most load-bearing unproven premise is the softmax-dominance condition, and the claimed open-world generalization is unsupported unless this condition is validated on truly unseen categories.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes 3D-CDRGP, a training-free policy that adapts closed-set clustering-based 3D object detectors to cross-device robotic grasping in open-world scenes. The method introduces SSGC-Seg, which converts point-wise semantic softmax scores into a binary foreground/background prediction using Eqs. (1)-(2), then clusters foreground points in the original coordinate system and scores proposals with ScoreNet-dagger, a modified scoring function that combines the backbone confidence with point count and mean height. The authors evaluate the approach on the Boxes dataset using PointGroup and PBNet as base detectors, and on two real robot systems (Elite EC66/AiNSTEC PRO and Dobo CR/RealSense D455) across five scenarios, reporting high recognition and grasping rates. The paper claims plug-and-play open-category grasping across devices without retraining or fine-tuning.","tokens_in":12763,"tokens_out":5242,"duration_ms":43042,"significance":"If the central claims hold, the paper offers a practical and deceptively simple solution to a real deployment problem: reusing a frozen closed-set 3D detector on new camera/robot hardware for category-agnostic grasping. The approach is compatible with multiple clustering-based detectors, and the authors demonstrate it on two genuinely different hardware systems with realistic objects, which is more than many grasping papers do. The decision to rely on geometry (point counts, height) rather than learned features for scoring is a sensible heuristic that is easy to reproduce. However, the evidence as presented is not yet sufficient: the key softmax-dominance premise behind SSGC-Seg is unvalidated on truly unseen categories, the real-world evaluation has very small implicit trial counts and no baselines on the same hardware, and the four manually chosen parameters are not ablated. The paper's value will be much stronger if these gaps are closed; as it stands, the central claim is plausible but under-supported.","major_comments":[{"comment":"The open-world claim rests on the assumption that for every foreground point on an unseen object, max(Si[2:M]) > Si[1], and for background points the reverse. The paper only states 'we observe' this behavior; it provides no quantitative evidence, no per-point foreground/background accuracy, and no analysis of how the softmax margin behaves as a function of object category, sensor, or distance. On the Boxes benchmark (Tab. 1) all objects are from the closed-set pre-training categories, so the rule's good performance there does not validate open-world operation. Since a violation of this assumption assigns an entire object to background (no proposal, no grasp) or table points to foreground (false proposals), the central 'open-category, cross-device' claim is currently unsupported. Please add a direct validation on held-out unseen categories, including per-point segmentation metrics and failure cases, or otherwise bound the condition.","section":"SSGC-Seg, Eqs. (1)-(2)"},{"comment":"The real-world evaluation reports only aggregate recognition and grasping rates, with no trial counts, no repeated runs, and no per-object breakdown. The single failure in Scenario 4 (spray can slipping) implies a denominator of nine trials (8/9 = 88.9%), which is too small to support the 100% success claims elsewhere. Moreover, there is no comparison on the same hardware against an alternative policy (e.g., the original ScoreNet, a closed-set baseline, or a generic instance-segmentation grasping method), so it is unclear whether the proposed SSGC-Seg and ScoreNet-dagger are responsible for the results. Please report trial counts, standard deviations or at least raw per-trial outcomes, and a same-hardware baseline comparison.","section":"Table 2 and 'Cross-Device Robotic Grasping Comparison'"},{"comment":"The pipeline has four manually set thresholds/weights (alpha=0.3, N_theta=60, C_theta=0.5, d_theta=2 in the Experiments section), and the reported success rates are conditional on these values. There is no ablation or sensitivity analysis showing how recognition/grasping rates vary with these parameters, nor any argument that the chosen values are robust across the two systems. If the method is intended as a plug-and-play policy, the parameter robustness should be demonstrated. Please add an ablation study, or at least sensitivity curves over plausible ranges for each parameter.","section":"Eq. (4) and Algorithm 1"}],"minor_comments":[{"comment":"The label 'SSGI-Seg' in the figure appears to be a typo for 'SSGC-Seg'; also, the arrow labels Fb, N_I^map, Hm are not all defined in the figure caption.","section":"Figure 2"},{"comment":"The phrase 'the pre-training model of he closed-set' should read 'the pre-training model of the closed-set'.","section":"Our Policy, Overview"},{"comment":"The notation yxmax and xymax is not defined clearly; the text should spell out that yxmax is the y-coordinate of the point with maximum x, and xymax is the x-coordinate of the point with maximum y.","section":"Eq. (6)"},{"comment":"The threshold C_theta is used in line 13 but is not defined in the 'ScoreNet-dagger' section; please define it where the other parameters are introduced.","section":"Algorithm 1"},{"comment":"The Boxes rows report only the proposed 'PointGroup + Ours' and 'PBNet + Ours' without baseline results on the same dataset; please include the unmodified closed-set detectors for reference.","section":"Table 1"},{"comment":"The phrase 'data collection from 2th to Mth index' is awkward; consider rewriting as 'the slice of S from index 2 to M'.","section":"SSGC-Seg section"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a robotics venue and addresses a practical deployment problem. The main concerns are the unvalidated softmax-dominance assumption and the thin real-world evaluation; I believe the authors can address these with additional experiments and reporting. I did not see evidence of citation manipulation, but the reference list includes a large number of self-citations and very recent preprints; the authors should double-check that all cited works are relevant and properly marked. The promise of code release should be enforced at acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThis paper is worth a serious look if you work on robotic grasping or open-world 3D detection. The core idea is simple: take a closed-set 3D instance segmentation model, treat “background” as the class whose softmax score beats every seen object class, and then cluster the remaining points geometrically to get graspable object proposals. They add a modified scoring function that prefers larger, higher objects, and they show the whole pipeline running on two different camera/robot combinations without retraining. The real-robot demos are genuine and the reported recognition/grasp rates are high. The cross-device angle is genuinely new — most prior work trains on augmented point clouds or adapts per hardware, and a training-free transfer is a useful thing to demonstrate.\n\nThe paper does several things well. The SSGC-Seg transformation (Eq. 1–2) is a neat adaptation of a closed-set detector to category-agnostic detection, and the authors are honest that it rests on an empirical observation about softmax behavior rather than a proof. The comparison of two clustering methods on the Boxes dataset is fine, and PBNet’s binary clustering being better than PointGroup’s distance clustering is a plausible finding. The writing is clear, and the supplementary videos (at least as described) back up the numbers.\n\nThe soft spots are real. First, the central assumption — that the background class always dominates for background points and that unseen foreground objects will trigger a seen-class softmax above background — is not stress-tested. The closed-set Boxes benchmark uses the same rectangular block categories that the model was pretrained on, so it cannot validate open-world behavior. The real-world scenarios use only a handful of objects, with no trial counts, no per-scene variance, and no failure analysis beyond one smooth spray can slipping. Second, there are four hand-set thresholds (alpha, N_theta, C_theta, d_theta) whose sensitivity is never examined; the high success rates are therefore partly fitted to the setups. Third, there are no baselines on the real systems — no closed-set detector without SSGC-Seg, no existing open-set detector, no retrained alternative. That makes it hard to know how much of the success comes from the specific SSGC-Seg trick versus the geometric clustering and the top-down grasping prior.\n\nNone of this makes the paper a reject in my view. The demonstration is real, the method is coherent, and the limitation section acknowledges the main gaps. But the open-world generalization claim is not yet fully supported. The authors should release code and data, report trial counts and variance, run at least one baseline on the same hardware, and analyze threshold sensitivity and the softmax-dominance condition on truly unseen categories.\n\nI’d send it to review, with the expectation of major revision. For you as a reader: worth a look for the idea, but do not treat the 100% numbers as established.","headline":"A genuine training-free cross-device grasping demo built on a softmax-background heuristic that is plausible but under-validated; deserves referee time but not yet a trusted open-world claim.","tokens_in":13371,"tokens_out":2607,"would_cite":false,"duration_ms":23615,"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 frozen 3D detector, with no retraining, can grasp unseen objects on new camera-and-robot hardware, the paper claims, and two real cross-manufacturer systems back it up.","keywords":["cross-device robotic grasping","3D open world","category-agnostic object detection","point cloud clustering","closed-set 3D detection","SoftMax foreground-background split","grasping without retraining"],"falsifier":"Run the frozen detector through SSGC-Seg on a scene with an unseen object that is similar in shape to a training class but placed on a cluttered or textured background (for instance, a doll resting on a patterned cloth), and measure the fraction of foreground points for which the original multi-class SoftMax assigns the top value to the background class; if that fraction is not near one, the proposal is fragmented and the reported near-100% grasp rates would not transfer to that scene.","tokens_in":12310,"feed_emoji":"🤖","tokens_out":8118,"duration_ms":61950,"temperature":0.7,"pith_summary":"This paper is trying to establish that a 3D object detector trained on a closed set of categories can be transplanted onto entirely different camera-and-robot systems and used to grasp objects it has never seen, with no retraining or fine-tuning. If true, a single pretrained detector can serve as a plug-and-play perception module for robotic picking across hardware generations, manufacturers, and object types. The policy replaces the detector's multi-class semantic output with a binary foreground/background split (using the observed tendency of the background class to dominate the SoftMax), clusters foreground points by geometry, and scores proposals with a revised rule that includes point count and height. The authors report near-perfect recognition and grasp rates on two real systems built from different vendors' cameras and arms, including stacked and everyday objects absent from the training set.","feed_headline":"No retraining: 3D detector grasps unseen objects on new robots","feed_subtitle":"A SoftMax foreground/background split lets a pretrained closed-set detector transfer across camera and robot brands.","key_machinery":"SSGC-Seg is the load-bearing module: it turns the multi-class semantic SoftMax output of a frozen closed-set detector into a binary foreground/background decision via Eqs. (1)–(2), exploiting the observed dominance of the background class, and then clusters the foreground points with the parameter-free geo-clustering of PBNet in the original coordinate system to form category-agnostic object proposals. ScoreNet‡ revises each proposal's confidence by combining the learned backbone feature score with the proposal's point count and mean height (Eq. 4), discarding proposals below a point-count threshold and guiding the gripper toward higher objects. Hand-eye calibration (Eq. 5) and a yaw-angle rule (Eq. 6) convert the proposals into robot grasps.","core_discovery":"The central claim is that the gap between closed-set 3D detection and open-world cross-device grasping can be closed without learning. The paper shows that a pretrained semantic branch, trained on a closed set of object classes, still behaves as a reliable background detector because the background class consistently receives the highest SoftMax value. Converting the multi-class SoftMax into a binary foreground/background score — the background score versus the maximum over all non-background classes — and clustering the resulting foreground points by geometry yields object proposals for arbitrary object categories. A revised score function, ScoreNet‡, adds point count and mean height to the learned confidence score, filtering out fragmentary false detections and prioritizing taller objects for top-down grasping. Across two complete robot–camera systems from different manufacturers (an AiNSTEC PRO camera with an Elite EC66 arm, and a RealSense D455 with a Dobo CR robot), the policy grasps previously unseen and stacked objects without retraining.","pith_inferences":["The background-dominance observation is a property of the specific pretrained model and its training distribution; the paper does not test whether the same dominance holds for models trained on other closed sets, so the policy's open-world generality across arbitrary detectors remains an open question.","A testable extension is to replace the hard argmax split with an uncertainty-aware criterion such as SoftMax entropy, which might preserve the gain when the background score is not strictly maximal.","Because deployment is training-free, the pipeline could be paired with vision-language models or open-vocabulary detectors to add object selection, turning open-category grasping into open-instruction grasping; the authors mention this as future work.","Clustering in original coordinates avoids offset-branch errors under domain shift but also discards learned object-center priors; on heavily occluded or interlocked scenes, that trade-off could reverse."],"forward_implications":["A pretrained closed-set 3D detector can be deployed on new camera hardware — different brands, resolutions, depth accuracies — with no data collection or fine-tuning.","The grasping system extends to arbitrary unseen object categories, not just training classes, as long as objects are separable from the background by geometry.","The policy is compatible with most clustering-based 3D instance-segmentation methods, so a stronger detector can be swapped in without changing the grasping pipeline.","Deployment cost drops: new robot arms and cameras only require kinematic setup and hand-eye calibration, while the perception model stays frozen.","The revised scoring rule (point count plus height) also suppresses fragmentary false detections and improves ordering in stacked scenes."],"supporting_citations":[{"why":"Supplies the canonical clustering-based 3D instance-segmentation pipeline (PointGroup) with semantic and offset branches, which the proposed policy adapts for open-world grasping.","marker":"(Jiang et al. 2020)"},{"why":"Provides PBNet, the binary-clustering detector whose geo-clustering step forms the foreground proposals and whose ScoreNet is revised into ScoreNet‡.","marker":"(Zhao et al. 2023)"},{"why":"Cited as the source of the MinkUnet sparse-convolution backbone that extracts point-cloud features for the pretrained detector.","marker":"(Li et al. 2019)"},{"why":"Cited as the sparse 3D convolution operation underlying the backbone feature extraction.","marker":"(Graham 2015)"}],"fun_headline_variants":["SoftMax split enables cross-device grasping without retraining","Open-world grasping on new robots: a SoftMax split","SoftMax cue transfers 3D grasp policy across devices","Grasp unseen objects across brands with a SoftMax split"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the pretrained semantic branch always gives the background class the highest SoftMax value, so that the binary-by-SoftMax split of Eqs. (1)–(2) cleanly separates graspable foreground points from the tabletop; if an unseen object makes the network spread probability across known classes or weakens the background's dominance, the proposals fragment and the grasp sequence misses targets.","fun_headline_variants_meta":{"raw":{"variants":["SoftMax split enables cross-device grasping without retraining","Open-world grasping on new robots: a SoftMax split","SoftMax cue transfers 3D grasp policy across devices","Grasp unseen objects across brands with a SoftMax split"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001292,"raw_usage":{"total_tokens":5294,"prompt_tokens":983,"completion_tokens":4311,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":599,"completion_tokens_details":{"reasoning_tokens":4243}},"tokens_in":599,"tokens_out":4311,"duration_ms":31172,"temperature":1.0,"reasoning_tokens":4243,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:27:55.089615+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the frozen detector through SSGC-Seg on a scene with an unseen object that is similar in shape to a training class but placed on a cluttered or textured background (for instance, a doll resting on a patterned cloth), and measure the fraction of foreground points for which the original multi-class SoftMax assigns the top value to the background class; if that fraction is not near one, the proposal is fragmented and the reported near-100% grasp rates would not transfer to that scene.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the canonical clustering-based 3D instance-segmentation pipeline (PointGroup) with semantic and offset branches, which the proposed policy adapts for open-world grasping."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides PBNet, the binary-clustering detector whose geo-clustering step forms the foreground proposals and whose ScoreNet is revised into ScoreNet‡."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Cited as the sparse 3D convolution operation underlying the backbone feature extraction."}],"review_version":1}