{"id":"0d61dd5f-bf4d-490d-a21f-7dcf96cb1c7f","arxiv_id":"2506.10818","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"high","formal_verification":"none","parameter_count":11,"one_line_summary":"An LSTM fed with fingertip-polygon features predicts time-to-grasp within about 21 ms, distance within 1 cm, and object size with over 97% accuracy in 4-fold tests, but leave-one-user-out errors are much larger.","lead":"This paper trains an LSTM network on fingertip motion to predict when a user will grasp an object, how far away the object is, and what size it is, using only hand dynamics during the reach. The authors report sub-21 ms timing errors and over 97% size accuracy for known users, but generalization to new users is much weaker.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline precision figures rest on 4-fold CV with overlapping sliding windows; adjacent windows share 24/25 samples, so train and validation are near-duplicates, and the paper's own L1UO results collapse to several-fold worse errors.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the random 4-fold split of overlapping subsequences creates near-duplicate windows across folds, so the reported precision figures reflect interpolation of the same reach rather than prediction of unseen reaches. This concern is concrete and located in the manuscript: Section 5 describes adding overlapping subsequences and then splitting into four equal parts without grouping by trial or participant. The paper's own leave-one-user-out results in Section 5.3 provide a strong independent check, and they show several-fold degradation (distance MAE 8.84 mm to 24.89 mm, time MAE 21.53 ms to 71.78 ms). That discrepancy makes the abstract's unqualified claims unsupported. I do not see a deeper internal inconsistency: the feature design, data collection, and LSTM architecture are reasonable, and the size prediction appears to survive user holdout better. The verdict of REJECT with moderate confidence is appropriate if the evaluation is not corrected; a trial-disjoint reanalysis could strengthen the paper, but as written the central quantitative claims are not established.","tokens_in":15101,"tokens_out":3562,"duration_ms":42702,"concrete_test":"Re-run the Section 5 pipeline exactly, but replace the random 4-fold split with trial-disjoint folds: assign every overlapping window extracted from the same participant-object-repetition R2G motion to the same fold, then compare mean MAE and accuracy for distance, time, and size. If these metrics stay near 8.84 mm, 21.5 ms, and 97%, leakage is not the driver; if they move toward the L1UO values of 24.89 mm, 71.78 ms, and lower accuracy, the abstract and discussion should be reworded to report known-user or calibrated performance and to emphasize the need for user adaptation. Ideally, release the split indices and code so the check is reproducible.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline claims are supported only by the 4-fold CV in Sections 5.1 and 5.2, and that protocol leaks labels. In Section 5, R2G sequences are chopped into fixed-length subsequences of 25, 50, or 75 samples, and overlapping subsequences are added to reach about 35,000 points; then the dataset is split into four equal parts. No grouping by participant or by original reach-to-grasp trial is described. At 960 Hz, 25 samples span roughly 26 ms, so consecutive windows differ by one sample and are near-identical. A random four-way split therefore places effectively the same window in training and validation, letting the LSTM memorize the local trajectory rather than predict unseen reaches. This is not a statistical quibble: when the user is held out in Section 5.3, distance MAE rises from 8.84 mm to 24.89 mm and time-to-grasp MAE from 21.53 ms to 71.78 ms (Figures 12 and 13), with object and size classification also degrading substantially. The abstract's unqualified '21 ms / 1 cm / 97%' numbers are thus not established for unknown reaches or users. The existence of some predictive signal is plausible, and the L1UO results remain better than chance, but the central quantitative claims as stated rest on the leaky 4-fold evaluation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a real-time grasp-prediction system based on local finger-motion dynamics. Using electromagnetic tracking of 16 participants performing 763 reach-to-grasp sequences at 960 Hz, the authors extract velocity and finger-polygon features and feed fixed-length windows into LSTM networks. The trained networks are used for three tasks: regression of the current distance to the target object, regression of the time until grasp, and classification of the object's size and shape. The headline results are mean MAEs of 8.84 mm for distance and 21.53 ms for time-to-grasp, with object-size classification accuracy above 97% under a four-fold cross-validation. Leave-one-user-out results are substantially worse (distance MAE 24.89 mm, time MAE 71.78 ms), with object discrimination often unusable before user-specific adaptation.","tokens_in":15445,"tokens_out":4783,"duration_ms":59927,"significance":"If the quantitative claims hold, this would be a useful contribution to HCI and mixed-reality interaction: local hand kinematics would enable early, high-precision prediction of grasp contact time and target identity, supporting latency mitigation and adaptive interfaces. The paper has clear strengths: a genuinely collected high-frequency hand-motion dataset (16 participants, 763 trials), a well-motivated feature set based on hand-polygon geometry, and an honest report of leave-one-user-out and transfer-learning results. The L1UO results also indicate that a real predictive signal exists even for unseen users, which gives the work scientific value beyond the headline numbers. The central weakness is that the headline evaluation protocol is contaminated by overlapping subsequences, so the specific precision figures stated in the abstract are not established by the paper as written.","major_comments":[{"comment":"The 4-fold cross-validation used for all headline numbers is leaky. The paper states that overlapping subsequences were added to balance the data and that the resulting dataset was then split into four equal parts, with no grouping by original R2G trial or participant. For the best-performing 25-sample sequence length at 960 Hz, consecutive windows differ by only one sample and share 24 of 25 samples, so a random four-way split places near-duplicate windows in both training and validation. The LSTM can therefore memorize local trajectory snippets rather than learn to predict unseen reaches. This directly affects the 8.84 mm distance MAE, the 21.53 ms time-to-grasp MAE, and the >97% size/object classification figures in Sections 5.1 and 5.2. The paper's own leave-one-user-out results (Section 5.3) show the magnitude of the effect: distance MAE rises to 24.89 mm, time MAE to 71.78 ms, and object discrimination becomes barely usable without user-specific adaptation. The abstract's unqualified '21 ms / 1 cm / 97%' claims are therefore not supported for unseen reaches or unseen users. The authors should redo the evaluation with splits that respect original trial and participant boundaries and report both within-user and across-user results in the abstract and conclusions.","section":null},{"comment":"The time-course analyses in Figure 8 and Figure 11 inherit the same leakage. The paper states that the network was trained on 75% of the 25-sample sequences and that run-time input was simulated with the remaining data; because those sequences overlap across the split, the claims that mean distance error falls below 15 mm already 400 ms before grasp and that accuracy remains above 95% are not trustworthy. These figures are used to support the central practical message that predictions are accurate well before contact. They should be recomputed on a held-out set constructed without overlapping windows or, better, on entire held-out reaches.","section":null},{"comment":"The abstract and the discussion present the 4-fold results as the general performance of the system, but the paper text itself in Section 5.3 concludes that object discrimination is 'barely usable without adaptation' for unknown users and that transfer learning with 150 user-dependent data points is needed to reach 87% (real) and 75% (synthetic) accuracy. This internal contradiction is not merely a presentation issue: the unqualified numerical claims in the abstract are misleading with respect to the paper's own more careful results. The authors should rephrase the abstract so that the high-precision claims are explicitly scoped to conditions where user-specific data are available, and they should state the leave-one-user-out figures as the primary generalization results.","section":null}],"minor_comments":[{"comment":"The text says 'word-space coordinate system' where 'world-space' is presumably intended; this typo should be corrected.","section":null},{"comment":"The figure captions contain 'Comparision' instead of 'Comparison'; also, the captions do not explain what the left and right panels represent as clearly as the text does.","section":null},{"comment":"The authors state that they verified differences between feature sets with inferential statistics but omit the details 'for clarity.' Given that the paper already reports standard deviations over folds, a brief note of the test used (e.g., paired t-test or Wilcoxon) and the significance level would improve reproducibility.","section":null},{"comment":"The hyperparameter selection is described only as progressively increasing the number of neurons until good performance is found. A short table listing the explored values and the final choices for each task would make the architecture section more complete.","section":null},{"comment":"The paper does not mention whether the data or source code will be made available. Given the paper's reliance on a newly collected dataset, a data-availability statement would strengthen the contribution.","section":null}],"recommendation":"major_revision","confidential_remarks":"I am not recommending outright rejection because the dataset and the qualitative phenomenon appear real, and the leave-one-user-out results indicate that the networks do extract a usable signal even before user calibration. However, the main quantitative claims as currently stated are invalidated by the overlapping-subsequence evaluation. If the authors cannot provide non-leaky cross-validation results that support the abstract's numbers, the manuscript should not be accepted. The paper would also benefit from a clearer separation between known-user, unknown-user, and after-adaptation performance in both the abstract and the conclusions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I'll be direct: the headline numbers are not trustworthy. The 4-fold cross-validation in Sections 5.1 and 5.2 splits overlapping subsequences from the same reach-to-grasp trials across train and validation. At 960 Hz, adjacent 25-sample windows differ by one sample, so the model is seeing near-duplicate inputs. The 21 ms, 1 cm, and 97% figures are therefore artificially optimistic.\n\nThat said, the paper has real merit. The finger-polygon feature set is a thoughtful way to capture pre-shaping without global position, and the data collection is careful: 16 participants, 960 Hz tracking, touch sensors for ground truth. The authors also run leave-one-user-out and transfer learning experiments, which show they are not hiding weaknesses. Those results are the honest story: distance MAE rises to 24.89 mm and time MAE to 71.78 ms, while object discrimination without calibration is poor. Size prediction holds up better under user holdout, which is genuinely interesting.\n\nThe main soft spot is the evaluation protocol. Overlapping windows are a known leakage source, and the abstract and discussion present the 4-fold numbers without qualification. The L1UO results should be the primary ones, or at least the headline should be scaled back. There is also no data or code release, which makes it hard to verify or build on the work. The hyperparameter search is described vaguely, but that is minor compared with the leak.\n\nThis paper is worth a serious peer review, not because it is close to acceptable as is, but because the question is relevant and the authors have done substantial groundwork. A major revision that redoes the analysis with trial- or participant-disjoint folds and reports true generalization numbers could turn this into a solid HCI contribution. I would accept it for review with a clear expectation of that revision.","headline":"A promising grasp-prediction idea undercut by a leaky cross-validation; the honest leave-one-user-out numbers tell a weaker but still interesting story.","tokens_in":15968,"tokens_out":3262,"would_cite":false,"duration_ms":36220,"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":"Local finger motion during a reach predicts grasp time to about 21 ms, distance to about 9 mm, and object size to over 97% accuracy.","keywords":["datasets","grasp prediction","hand gesture","neural networks","LSTM","reach-to-grasp","hand kinematics","finger motion dynamics"],"falsifier":"Re-run the evaluation with data split by entire reaches and then by participant, so no window from a training motion appears in the validation set; if the time-to-grasp error rises substantially above the reported roughly 21 ms, the claim of high-precision prediction for known users must be scaled back.","tokens_in":14904,"feed_emoji":"🖐️","tokens_out":7547,"duration_ms":80020,"temperature":0.7,"pith_summary":"This paper tries to establish that the hand's local finger dynamics during a reach-to-grasp movement contain enough information to anticipate the imminent grasp before the hand arrives. On a dataset of 763 reach-to-grasp sequences from 16 participants grasping 16 real and synthetic objects, the authors trained small LSTM networks on short windows of hand-polygon features and hand speed. They report a mean absolute error of about 21.5 ms for the remaining time until grasp, about 8.8 mm for the current distance to the target, and better than 97% accuracy for target size in their main cross-validation. The practical payoff is an interactive system that can react before contact, shrinking perceived latency and enabling predictive haptics, object disambiguation, and real-time hand-redirection models.","feed_headline":"Fingertip motion predicts grasp time to under 21 ms","feed_subtitle":"A simple LSTM reads hand-polygon features to anticipate when and what a user will grab before contact.","key_machinery":"The hand polygon model, built from the five fingertip sensors, represents the grip by a set of edge vectors that encode thumb-index aperture, thumb-little-finger aperture, curvature, and hand orientation while subtracting out global hand position. These 15-dimensional fingertip-polygon features, optionally augmented by a proximal-phalanx polygon, are concatenated with hand speed into 25-sample windows (about 26 ms at the 960 Hz capture rate) and fed to an LSTM with a fully connected output head. The local, hand-relative nature of the features is what allows the network to predict grasp time, distance, and size from a short slice of the motion.","core_discovery":"The central claim is that the relative configuration of the fingertips, encoded as the hand polygon formed by five fingertip positions plus hand speed with global hand position removed, is sufficient for real-time prediction of grasp-related quantities without instrumenting the target object. A single-layer LSTM with 64 hidden units regresses both time-to-grasp and distance-to-target from a 25-sample window, and adding proximal-phalange positions does not improve the result. The same architecture classifies target size and, for objects with distinct grasp affordances, the target identity, with mean discrimination accuracy above 95% during the last 400 ms of the reach. The authors are careful to report that shape discrimination among deliberately similar synthetic solids is much harder, and that when users are held out entirely, distance error rises to about 25 mm and time error to about 72 ms until the network is fine-tuned with a small amount of the new user's data.","pith_inferences":["Because the 4-fold validation splits overlapping 25-sample windows rather than whole trials, the headline numbers should be read as upper bounds; the leave-one-user-out results, about 25 mm distance and 72 ms time before adaptation, are the more honest estimate for a first-time user.","A natural next experiment is to close the loop in a real latency-sensitive task, measuring whether acting on the predicted grasp time actually removes perceived latency rather than only reporting offline error.","The size classifier could be replaced by a continuous grasp-aperture regressor, which would let the same features predict the size of objects never seen in training.","Fusing gaze direction with the hand-polygon features may resolve the shape confusions the paper documents, since gaze is an independent signal for the intended target during the reach."],"forward_implications":["Hand-redirection and reach-modeling techniques that currently normalize the whole reach to a unit interval can be driven in real time, because the predicted time-to-grasp supplies the missing time-scaling parameter.","Interfaces can start rendering or executing an action roughly 280–400 ms before contact, directly offsetting point-to-point latency.","Fingertip-only tracking hardware, such as lightweight magnetic or ring-based sensors, is enough to obtain the predictions, so no instrumented glove or global hand pose is required.","Distance and size predictions together can narrow the set of candidate targets in mixed reality, especially when the environment is designed with distinguishable grasp affordances.","Object identity prediction is only reliable when objects differ in grasp affordances; visually or grasply similar objects will be confused, so the method suits affordance-aware interface design more than general object recognition."],"supporting_citations":[{"why":"shows hand kinematics carry information about the size of the intended object, the perceptual basis the paper exploits","marker":"[4]"},{"why":"reviews evidence that grasp formation correlates with object form, size, and intended action","marker":"[8]"},{"why":"demonstrates that fingertip positions can be tracked precisely with low-cost magnetic sensing, supporting the lightweight front end","marker":"[14]"},{"why":"shows hand pre-shaping is largely finalized before the hand reaches the object, the temporal window the predictions depend on","marker":"[18]"},{"why":"provides the reach-prediction baseline for haptic retargeting that this work extends with finger dynamics","marker":"[19]"},{"why":"shows grasp recognition from uncalibrated glove data, a comparable HCI baseline for object discrimination","marker":"[34]"},{"why":"documents gradual molding of the hand to object contours, grounding early object-size readout","marker":"[46]"},{"why":"is the grasp-posture object recognizer whose 60% general and 98% user-specific recognition rates frame the comparison","marker":"[48]"}],"fun_headline_variants":["Fingertip geometry predicts grasp time to 21 ms","Hand polygon anticipates grabs within 21 ms","LSTM reads hand shape to forecast grasp distance","Finger motion reveals object size with 97% accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The headline accuracy numbers assume that overlapping 25-sample windows cut from the same reach-to-grasp motion can be treated as independent training and test examples, so the reported precision may be artificially high.","fun_headline_variants_meta":{"raw":{"variants":["Fingertip geometry predicts grasp time to 21 ms","Hand polygon anticipates grabs within 21 ms","LSTM reads hand shape to forecast grasp distance","Finger motion reveals object size with 97% accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000376,"raw_usage":{"total_tokens":1968,"prompt_tokens":872,"completion_tokens":1096,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":488,"completion_tokens_details":{"reasoning_tokens":1033}},"tokens_in":488,"tokens_out":1096,"duration_ms":10014,"temperature":1.0,"reasoning_tokens":1033,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:17:24.494637+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the evaluation with data split by entire reaches and then by participant, so no window from a training motion appears in the validation set; if the time-to-grasp error rises substantially above the reported roughly 21 ms, the claim of high-precision prediction for known users must be scaled back.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"reviews evidence that grasp formation correlates with object form, size, and intended action"},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"shows hand pre-shaping is largely finalized before the hand reaches the object, the temporal window the predictions depend on"},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"shows grasp recognition from uncalibrated glove data, a comparable HCI baseline for object discrimination"},{"cited_title":"Santello and J","cited_arxiv_id":null,"evidence_quote":"documents gradual molding of the hand to object contours, grounding early object-size readout"}],"review_version":1}