{"id":"4d836a9f-0051-4b76-977a-5cc0577ce057","arxiv_id":"2501.06987","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A simulation-based pipeline using GraspIt grasp quality metrics detects hand-object contact with 89.3% accuracy on DexYCB frames.","lead":"This paper builds a system that detects when a human hand is touching an object by reconstructing the grasp in a physics simulator and reading its grasp-quality scores. It reports about 89% agreement with manually labeled grasp frames in the DexYCB dataset, a proof of concept for sensor-free contact detection in robot handovers.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported 89.3% accuracy rests on undocumented manual 'stable grasp' labels that the authors themselves call error-prone for subjects 8 and 9; because the detector flags contact while the labels mark stable grasp, the headline number may partly measure annotation artifacts rather than contact…","rationale":"I read the paper as a proof-of-concept: reconstruct the MANO hand and YCB object in GraspIt! using DexYCB ground-truth poses, declare contact when either Ferrari-Canny metric is nonzero, and compare against manually labeled stable-grasp frames. The internal binary behavior of the metrics is plausible and likely reproducible, so I do not see a formal contradiction in the simulator step itself. The insecure condition is the reference: the manual stable-grasp label is simultaneously the target of comparison and the source of acknowledged error. The reader's weakest_assumption identifies the same issue. My proposed check re-annotates a sample with a written protocol and consensus labels; it separates 'contact' from 'stable grasp' and would show whether 89.3% survives annotation noise. If it survives, the conditional verdict stands as an acceptable proof-of-concept; if not, the central accuracy claim needs to be downgraded or the annotation protocol supplied. I am not raising a separate baseline objection here, though a contact-flag baseline would be a useful additional control; the most load-bearing risk remains the label construct.","tokens_in":4981,"tokens_out":7221,"duration_ms":79535,"concrete_test":"Take a random stratified sample of about 500 frames from the 22,659 evaluated frames, covering all seven subjects and all twenty objects. Have two independent annotators label each frame under a written protocol that distinguishes (i) any visible hand-object contact and (ii) a stable grasp. Compute Cohen's kappa for each label type, then recompute the overall, per-object, and per-subject accuracy using only frames where consensus contact labels exist. If kappa is below 0.7, or if the recomputed overall accuracy differs from 89.3% by more than 2 percentage points, the headline accuracy is not robust to annotation noise and the central claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical claim is the 89.3% agreement between the system's contact detections and the ground truth. That ground truth is not contact annotation but manual labeling of frames in which a subject 'appeared to achieve a stable grasp' (Section III). Contact and stable grasp are different constructs: a hand can touch an object without achieving a stable grasp, and the authors themselves interpret early detections as contact before the stable grasp is confirmed, which would be scored as false positives. The paper gives no annotation protocol, no inter-annotator agreement, no count of frames deemed ambiguous or excluded, and no analysis of how label noise propagates into Table I and Table II. It even attributes the lower accuracy for subjects 8 and 9 to 'errors in our manual ground truth annotation process' (Section III.A), an admission that the reference labels are not a fixed standard. If those labels are noisy or systematically shifted in time, the 89.3% figure is not a clean measure of contact detection; it confounds detector error with label error. This is load-bearing because the entire evaluation and the paper's headline number depend on this comparison, rather than on any external sensor or repeatable criterion.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a hand-object contact detection system that reconstructs the human hand (MANO model) and the grasped object in the GraspIt! simulator, then uses the Ferrari-Canny grasp quality metrics epsilon and v to decide whether contact has occurred. The detection rule is simple: contact is declared if either metric is greater than zero, based on the observed simulator outputs of epsilon=-1.0 and v=0.0 for no-contact frames. The method is evaluated on the DexYCB dataset using ground-truth hand and object poses, and the authors report an overall accuracy of 89.3% against manual annotations of frames in which a subject appeared to achieve a stable grasp. Per-object and per-subject accuracy tables are provided, along with distribution plots of the quality metrics and a correlation analysis between detection timing and false-positive rate.","tokens_in":5216,"tokens_out":3388,"duration_ms":35583,"significance":"The central idea is interesting and potentially useful: using grasp-quality metrics as a contact indicator could enable contact detection without tactile or force sensors, which is relevant for robot-to-human handovers. The paper is honest about relying on idealized ground-truth poses and about the limitations of the manual annotation. If the evaluation were made rigorous, the approach could be a practical contribution. However, the current evidence does not yet support the headline accuracy claim, because the ground truth is not contact annotation, no baseline is compared, and no uncertainty measures are given. The paper does provide a reproducible pipeline description and uses a public dataset, which are strengths, but the assessment is not yet at the level required for a journal publication.","major_comments":[{"comment":"The ground-truth labels are manual annotations of frames where subjects 'appeared to achieve a stable grasp,' not annotations of hand-object contact. The system, by contrast, detects contact using grasp-quality metrics. Contact and stable grasp are different constructs: a hand can contact an object before a stable grasp is achieved, and the paper itself interprets early detections as false positives in Fig. 5. Therefore the reported 89.3% accuracy (Section III.A) is not a clean measure of contact detection; it measures agreement with a stable-grasp label whose temporal alignment is not defined. The authors must either use a contact ground truth (e.g., contact labels in DexYCB, if available) or explicitly model and report the expected mismatch. At minimum, the annotation protocol, inter-annotator agreement, and the number/type of excluded frames must be reported, especially because the authors attribute accuracy dips for subjects 8 and 9 to 'errors in our manual ground truth annotation process.' This issue is load-bearing because the entire evaluation depends on this comparison.","section":"Section III, Section III.A"},{"comment":"No baseline method is evaluated. The paper's central claim is that the system achieves high accuracy, but without comparing against a trivial geometric baseline (e.g., distance between hand and object meshes) or an existing contact-detection method, the 89.3% figure cannot be interpreted as an improvement or even as a calibrated result. The introduction contrasts force-based methods with the proposed approach, but no quantitative comparison to any alternative is provided. Adding at least one simple baseline is necessary to support the claim that grasp-quality metrics provide a useful contact signal beyond what a distance threshold would give.","section":"Section III.A, Tables I and II"},{"comment":"No statistical uncertainty or significance measures are reported. The per-object accuracy ranges from 83.9% to 93.2% and per-subject from 85.1% to 91.1%, but these are point estimates without confidence intervals, and the differences are discussed without any significance testing. Figure 5 is described as showing a 'correlation' between detection offset and false-positive rate, but no correlation coefficient, p-value, or goodness-of-fit is reported, and the regression line is presented with no error bounds. As a result, the reader cannot assess whether the observed patterns are meaningful or how stable the headline 89.3% accuracy is across the dataset.","section":"Section III.A and Figure 5"}],"minor_comments":[{"comment":"The text states that both quality metrics have 'roughly similar distributions regardless of ground truth,' which appears to undermine the discriminative power of the metrics; this tension should be discussed more explicitly, perhaps by quantifying the overlap or by showing that only the temporal change (Fig. 7) is informative.","section":"Section III.B"},{"comment":"The evaluation section mentions 'statistical analysis ... of processing times,' but no processing-time results appear anywhere in the paper; either report them or remove the mention.","section":"Section III"},{"comment":"The conversion from MANO parameters to the URDF robot description used by GraspIt! is described only briefly; a short explanation or a reference to the exact procedure would improve reproducibility.","section":"Section II"},{"comment":"The abstract and conclusion state accuracy 'approaching 90%' while the body reports 89.3%; for a paper that emphasizes numbers, the abstract should state the precise figure.","section":"Abstract and Conclusion"},{"comment":"Reference [11] is a URL with an access date; consider replacing it with a standard citation or adding a formal title.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper reads more like a workshop contribution than a full journal article: the evaluation is short, lacks baselines and uncertainty quantification, and the ground-truth definition is not aligned with the contact-detection task. The core idea is sound and could be published after a substantial revision that addresses the evaluation gaps. I would not reject outright, but the current form is not yet acceptable for a serious journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a clean little experiment that basically checks whether GraspIt's contact detector, thresholded by its Ferrari-Canny quality metrics, lines up with manual stable-grasp annotations in DexYCB. It mostly does, 89.3% overall. That's a useful sanity check, but the paper oversells it slightly by calling the manual labels 'ground truth' for contact.\n\nWhat's new: the specific pipeline — MANO hand reconstruction, GraspIt, Ferrari-Canny metrics, DexYCB — appears to be new. The authors are honest about the idealized setup (ground truth poses) and about their own annotation issues for subjects 8 and 9.\n\nWhat it does well: the evaluation is per-object and per-subject, the false positive / false negative breakdown is informative, and the time-series observation (metrics jump at the transition) is a genuine insight worth following up.\n\nSoft spots, in order of importance. First, the labels are 'appeared to achieve a stable grasp', not contact. A hand can touch an object without a stable grasp, and the authors even interpret early detections as contact before the stable grasp is confirmed. So the 89.3% is agreement with a stable-grasp heuristic, not a clean contact detection rate. Second, no annotation protocol, no inter-annotator agreement, no count of excluded or ambiguous frames. The authors' own comment about subjects 8 and 9 suggests label noise is real. Third, no baseline. A simple mesh-distance threshold might do as well or better; without it, we don't know what the quality metrics add over a geometric check. Fourth, no error bars; the per-object range (83.9–93.2) suggests real variation but we can't tell what's signal.\n\nThe core idea is sound and the paper is transparent. It's a proof-of-concept, not a definitive study. The numbers will hold up as a sanity check, but the central claim should be read as 'our thresholded simulator contact detector agrees with manual stable-grasp labels in most frames', not as a general contact detection accuracy.\n\nWho this is for: robotics people working on human-robot handover or contact detection without force sensors. It's a fine workshop/short paper. For a full journal, it needs a baseline, a proper annotation study, and ideally an error analysis separating label noise from detector error.\n\nI'd send it to peer review — it's honest and reproducible — but I'd push for major revision on the evaluation. If the authors add a baseline and tighten the label definition, it becomes a solid short paper.","headline":"A small, honest proof-of-concept: GraspIt's built-in contact state is used as a contact detector, and it agrees with manual stable-grasp labels ~89% of the time, but the evaluation lacks a baseline and the labels are not contact labels.","tokens_in":5735,"tokens_out":2205,"would_cite":false,"duration_ms":21418,"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":"Under ground-truth poses, simulator grasp-quality scores detect hand-object contact in 89.3% of frames.","keywords":["contact detection","grasp detection","grasp quality metrics","Ferrari-Canny quality measures","scene reconstruction","robot-to-human handover","MANO hand model","DexYCB"],"falsifier":"Run the system on frames with synchronized tactile or force ground truth instead of visual annotation: if a significant number of frames with physical hand-object contact but non-force-closure hand poses have $\\epsilon = -1.0$ and $v = 0.0$, then the method detects grasps rather than contact, and if many no-contact near-miss poses yield $\\epsilon > 0$ or $v > 0$, then the simulator's contact model overreports; either outcome would refute the claim that the metric criterion measures hand-object contact.","tokens_in":4785,"feed_emoji":"🤖","tokens_out":8815,"duration_ms":80492,"temperature":0.7,"pith_summary":"Under idealized conditions, with ground-truth hand and object poses fed directly into a grasp simulator, this paper tries to show that hand-object contact can be detected from grasp quality scores rather than from force or tactile sensors. Reconstructing the interaction scene and reading two Ferrari-Canny stability measures, the system declares contact whenever either measure rises above its no-contact baseline. Compared with manually annotated stable-grasp frames from a human-object interaction benchmark, this rule matches the human label in 89.3% of frames, with a 15.7% false-positive rate and an 8.0% false-negative rate over 22,659 frames. The authors argue this matters for robot-to-human handover, where deciding when a person has actually taken an object is safety-critical and force-based methods can release too early if the hand merely touches the object. It would also give a way to test whether a hand pose is graspable at all.","feed_headline":"Simulated grasp scores flag hand-object contact in 89.3% of frames","feed_subtitle":"A simulation-based check of hand and object poses could replace force sensors in robot-to-human handovers.","key_machinery":"The load-bearing object is the pair of Ferrari-Canny grasp quality measures $\\epsilon$ and $v$ computed by the GraspIt! simulator: $\\epsilon$ is the worst-case grasp stability and $v$ is the average-case stability. The simulator returns sentinel values $\\epsilon = -1.0$ and $v = 0.0$ when there is no contact, and at least one of the two becomes positive when contact is made, which the paper turns into the decision rule contact $\\Longleftrightarrow$ ($\\epsilon > 0$ or $v > 0$). The hand is represented with a MANO mesh (shape vector $\\beta$ and pose vector $\\theta$), converted into a joint-based kinematic model in the simulator, and the object's VRML model is placed according to its ground-truth pose. The machinery converts arbitrary hand and object poses into scalar stability signals without any learned classifier.","core_discovery":"The paper's central claim, stated in Section III.A, is that the system correctly detects hand-object contact in 89.3% of evaluated frames. The detection criterion is exactly that the simulator reports no-contact sentinel values ($\\epsilon = -1.0$ and $v = 0.0$) when there is no contact, and contact is asserted when $\\epsilon > 0$ or $v > 0$. Against manual labels of when a subject appears to achieve a stable grasp, this criterion yields object-wise accuracy from 83.9% (a bowl) to 93.2% (a pudding box) and subject-wise accuracy from 85.1% to 91.1%. The two grasp quality measures have overlapping distributions for grasped and non-grasped frames, so they are not cleanly separable by a threshold, but higher values are slightly more common in stable-grasp frames and both measures jump when the ground-truth state transitions into grasping.","pith_inferences":["If the manual labels encode 'stable grasp' rather than physical touch, then the reported 89.3% agreement conflates contact detection with grasp-stability classification; evaluating the same pipeline against synchronized tactile-sensor ground truth would likely show different accuracy on non-grasp touches.","The strong overlap in quality-metric distributions suggests that a simple threshold on $\\epsilon$ or $v$ will not separate contact from non-contact in general, so a temporal detector that scores the jump in both metrics at the grasping transition may outperform the static rule.","The method could be extended to estimate grasp readiness continuously during handover by treating the rising $\\epsilon$ and $v$ time series as evidence accumulation, and this could be tested without new hardware by replaying the benchmark sequences under different decision rules.","A testable implication of the false-positive correlation is that if the ground-truth labels were shifted earlier to the moment of first physical contact, the reported accuracy would rise; re-annotating a subset of frames with contact sensors would settle whether the offset is labeling lag or detector lead."],"forward_implications":["In a robot-to-human handover, object release can be triggered by the simulator's contact signal instead of by tactile sensors, potentially avoiding premature release when the human touches but does not securely grasp the object.","The same pipeline can flag whether a hand is currently in a graspable pose, extending the output from a binary contact label to a continuous stability estimate.","Because hand shape parameters are person-specific but estimable once per subject, a real-time vision-based version is plausible using a single RGB image for hand reconstruction and object pose estimation.","Accuracy depends on object geometry: small objects that are easily grasped on their sides, such as a pudding box, are detected best, while a bowl that can be pinched or cupped is detected worst, so the method is most trustworthy for power-grasped objects.","False positives correlate with the time between first detected contact and the manual stable-grasp label, meaning the system tends to report contact earlier than a human's 'stable grasp' annotation."],"supporting_citations":[{"why":"Supplies the GraspIt! simulator used to reconstruct the interaction scene and compute the grasp quality metrics.","marker":"[3]"},{"why":"Defines the Ferrari-Canny grasp quality measures $\\epsilon$ and $v$ that the contact criterion is built on.","marker":"[5]"},{"why":"Provides the MANO hand model whose shape and pose parameters reconstruct the human hand in simulation.","marker":"[4]"},{"why":"Is the benchmark dataset that supplies the ground-truth hand and object poses and the frames being evaluated.","marker":"[7]"},{"why":"Provides the YCB object set from which the manipulated objects' 3D models are taken.","marker":"[8]"}],"fun_headline_variants":["Grasp quality metrics detect hand-object contact at 89.3% accuracy","Contact detection from grasp scores: 89.3% accurate","Hand-object contact via grasp quality hits 89.3%","Grasp-based contact detection achieves 89.3% accuracy","Simulated grasp scores serve as contact detector at 89.3%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the manually annotated labels, marking frames where a subject appeared to achieve a stable grasp, are reliable enough to serve as ground truth; the paper does not describe the annotation protocol or inter-annotator agreement and itself attributes lower accuracy for two subjects to annotation errors.","fun_headline_variants_meta":{"raw":{"variants":["Grasp quality metrics detect hand-object contact at 89.3% accuracy","Contact detection from grasp scores: 89.3% accurate","Hand-object contact via grasp quality hits 89.3%","Grasp-based contact detection achieves 89.3% accuracy","Simulated grasp scores serve as contact detector at 89.3%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000193,"raw_usage":{"total_tokens":1267,"prompt_tokens":781,"completion_tokens":486,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":397,"completion_tokens_details":{"reasoning_tokens":393}},"tokens_in":397,"tokens_out":486,"duration_ms":4754,"temperature":1.0,"reasoning_tokens":393,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:48:41.614768+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the system on frames with synchronized tactile or force ground truth instead of visual annotation: if a significant number of frames with physical hand-object contact but non-force-closure hand poses have $\\epsilon = -1.0$ and $v = 0.0$, then the method detects grasps rather than contact, and if many no-contact near-miss poses yield $\\epsilon > 0$ or $v > 0$, then the simulator's contact model overreports; either outcome would refute the claim that the metric criterion measures hand-object contact.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the GraspIt! simulator used to reconstruct the interaction scene and compute the grasp quality metrics."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Ferrari-Canny grasp quality measures $\\epsilon$ and $v$ that the contact criterion is built on."},{"cited_title":"Romero, D","cited_arxiv_id":null,"evidence_quote":"Provides the MANO hand model whose shape and pose parameters reconstruct the human hand in simulation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Is the benchmark dataset that supplies the ground-truth hand and object poses and the frames being evaluated."},{"cited_title":"Calli et al., ”Yale-CMU-Berkeley dataset for robotic manipulation research,” The International Journal of Robotics Research, 2017","cited_arxiv_id":null,"evidence_quote":"Provides the YCB object set from which the manipulated objects' 3D models are taken."}],"review_version":1}