{"id":"23d0fc3c-ac62-4660-b593-b2b81745b036","arxiv_id":"2411.08499","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A tactile-driven policy that adapts gripper angle in real time improves the maximum disturbance a grasped object can withstand.","lead":"This paper trains a robot gripper to adjust its grip angle using tactile sensor feedback when a grasped object is pulled, shaken, or filled with water. The gripper resists more force with the learned adaptation than without it, across seven everyday objects.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'seven novel objects' generalization claim is not established because the paper never lists the 12 demonstration objects or states that the test objects are disjoint, and Table I lacks a control that rules out a trivial grip-tightening reflex.","rationale":"The reader identified tactile-state sufficiency as the weakest assumption. I do not think that is the most load-bearing issue: 32 tactile channels plus joint angle are a plausible sensor suite for detecting slip and force changes in a two-finger gripper, and the demonstration protocol appears to capture the relevant signals. The more serious gap is that the paper's central generalization and improvement claims are not backed by a controlled evaluation: no object split disclosure, no baseline comparisons, and no repeated trials. This is not an internal inconsistency in the method, but it is load-bearing because Table I is the only quantitative evidence for the contribution. If the seven test objects overlap with the 12 demonstration objects, the generalization claim collapses; if a fixed-Δθ controller matches the learned policy, the 'learning' contribution is not supported. Both are concrete and checkable. I therefore retain the reader's CONDITIONAL verdict, but with the condition refocused on object-level disjointness and baseline control rather than tactile-state sufficiency.","tokens_in":7429,"tokens_out":5078,"duration_ms":51205,"concrete_test":"Request the authors to release the identities of the 12 training objects and the object-level train/test split. If any Fig. 5 object appears in the demonstration set, re-run the Table I protocol on a strictly disjoint object set. Independently, run one control condition on the same seven objects in which the controller closes the gripper by a fixed Δθ (e.g., 2° or 5°) whenever the stability estimator flags instability, and compare the maximum supported weight to the learned he-AGPT policy. If the fixed-closing baseline matches the learned policy, the 'adaptation skill' claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central empirical claim—that the trained model 'generalizes to daily objects with seven different sizes, shapes, and textures' and that Table I's higher supported weights demonstrate the benefit of learned adaptation—depends on the seven test objects being genuinely novel relative to the model. Section III-B reports '1200 experiments were performed on 12 objects' for demonstration and training, but never identifies those 12 objects. Section III-D then calls the Fig. 5 objects 'novel' without stating that they are disjoint from the training set. If any of the seven test objects (pillbox, tea can, mouthwash, milk bottle, wine bottle, perfume, ink) were among the 12 demonstration objects, the claimed generalization is not demonstrated, and Table I could reflect object-specific memorized adjustments rather than a generalizable tactile policy. In addition, Table I provides no trial counts, error bars, or comparison against a non-learned baseline that simply closes the gripper by a fixed Δθ when instability is detected. Since the adaptation action α = Δθ changes gripper angle and therefore grip force, the reported improvement could be reproduced by a trivial force-increase reflex. This is the load-bearing weak point: the quantitative evidence does not yet rule out the alternatives of training-set contamination or a simple grip-tightening controller.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a tactile-driven adaptive grasping framework, called he-AGPT, consisting of three components: an initial grasp generator trained by behavior cloning, a Gaussian Mixture Model (GMM) stability estimator, and a self-attention-based grasp adaptation policy that outputs corrective gripper-angle changes. The system is evaluated on a real electric parallel gripper with custom 32-channel tactile sensors. The experimental section reports, for seven everyday objects, the maximum graspable weight with and without the adaptation strategy, and the authors claim that the learned policy generalizes to novel objects and resists disturbances such as human pulling, vibration, and added mass. The central claim is that tactile-only feedback plus gripper angle suffices to detect instability and generate corrective actions without vision or object models.","tokens_in":7720,"tokens_out":6082,"duration_ms":55864,"significance":"If the claims are established, the paper would provide a useful demonstration that a purely tactile, learned policy can maintain grasp stability on a parallel gripper under external disturbances. The use of real hardware with 1200 human demonstrations and a custom tactile sensor is a concrete strength, as is the clear division of the system into an initial grasp generator, a stability estimator, and an adaptation policy. However, the current evidence does not yet support the generalization claim: Table I reports no trial counts, variance, or statistical comparison; the training and test object sets are not shown to be disjoint; and the adaptation policy is not compared against a simple grip-tightening baseline. These are experimental-reporting gaps that could be addressed with additional experiments and clarifications, rather than fundamental flaws in the core idea.","major_comments":[{"comment":"The central generalization claim is not supported by the reported data. Section III-B states that 1200 experiments were performed on 12 objects for demonstration and training, but it never identifies those 12 objects, and Section III-D calls the objects in Fig. 5 'novel' without stating that they are disjoint from the training set. If any of the seven test objects (pillbox, tea can, mouthwash, milk bottle, wine bottle, perfume, ink) were among the 12 training objects, Table I could reflect object-specific memorized adjustments rather than a generalizable tactile policy. The authors should list the 12 training objects, state explicitly the overlap with the test set, and clarify how the 30,000 data points (10,000+6,000+14,000) are allocated among the three modules.","section":"Sec. III-B and III-D, Table I"},{"comment":"The quantitative evidence does not rule out a trivial grip-tightening reflex. Because the only adaptation action is alpha = Delta_theta (Eq. (7)), closing the gripper by a fixed amount whenever instability is detected would increase grip force and could plausibly reproduce the 'with adaptation' column in Table I. Table I reports no trial counts, no error bars, no confidence intervals, and no comparison with such a non-learned baseline. The claim that the learned policy, rather than the action space itself, is responsible for the observed improvements is therefore unsubstantiated. The authors should report repeated trials with variance and include a baseline that simply applies a fixed positive Delta_theta on instability detection.","section":"Table I and Sec. III-D"},{"comment":"The stability-threshold definition is mathematically malformed. Eq. (6) defines Lik_{2sigma}(i) as (2*pi)^{-d/2} |Sigma_i|^{-1/2} e^{-2}, which is not the Gaussian likelihood at two standard deviations; the exponent is an unexplained constant, the dimension d is not defined, and the relationship to 'two standard deviations' is unclear. Since a and b are then used to select the discrimination threshold t_e, and this threshold controls when adaptation is triggered, the formula must be corrected. The authors should also specify which dataset (training or validation) is used for ROC-based threshold selection.","section":"Sec. II-B, Eq. (6)"},{"comment":"The training objective is internally inconsistent. Eq. (1) defines the policy as pi_w(s), Eq. (2) updates parameters theta with pi_theta(s_i) without defining theta in this context, and Eq. (3) defines a mean-squared-error loss, whereas Section III-C states that the initial grasp generator uses cross-entropy loss with SGD. This makes it impossible to verify the training procedure. The authors need to align the notation across equations and state the actual loss function used for each network.","section":"Sec. II-A, Eqs. (1)-(3), and Sec. III-C"},{"comment":"The state definition is inconsistent with the policy input. The stability estimator defines the current grasp as X* = (S*, theta*, P*) with P in R^7 representing the end-effector location, and Eqs. (4)-(5) include P components in the GMM. However, the adaptation dataset in Eq. (9) contains only Delta_S, S, theta, and Delta_theta, with no P, and the self-attention input in Eq. (10) is Xi = (Delta_S_i, Delta_theta_i). If the end-effector pose is not observed by the adaptation policy, the definition in Eq. (4)-(5) is misleading; if it is observed, the policy input must be specified accordingly. This matters for the tactile-only claim, since the system must be shown to rely only on the information actually fed to the learned components.","section":"Sec. II-B and II-C"}],"minor_comments":[{"comment":"The abstract contains grammatical errors: 'an human-demonstration-based adaptive grasping policy base on tactile' should be 'a human-demonstration-based adaptive grasping policy based on tactile sensing'.","section":"Abstract"},{"comment":"The conclusion's admitted limitation, that the gripper shows limited adaptability for targets with abrupt shape changes, should be reconciled with the abstract's claim of 'excellent generalization ability'.","section":"Sec. IV"},{"comment":"The Fig. 4 caption contains 'with with adding water', and the Fig. 7 caption repeats 'A snapshot for the experiment on wine bottle' twice.","section":"Fig. 4 and Fig. 7 captions"},{"comment":"The sentence 'Once the stability proposed in III-B inform us the current grasp is unstable' should refer to Section II-B and should read 'informs us'.","section":"Sec. II-C"},{"comment":"The update rule uses the index i both for the dataset elements and for the gradient-step counter; use separate indices to avoid confusion.","section":"Eq. (2)"},{"comment":"For the GMM stability estimator, the text says 'the number of classes set to 2', which is ambiguous: it should clarify whether this is the number of mixture components m or the number of stability labels.","section":"Sec. III-C"}],"recommendation":"major_revision","confidential_remarks":"The manuscript appears to describe a working hardware system, but the evaluation is currently too weak for archival publication. The main gaps are the unspecified overlap between training and test objects, the absence of trial counts and variance, and the lack of a non-learned grip-tightening baseline. These are addressable with additional experiments and clarifications, so I recommend major revision rather than rejection. The paper also needs a careful proofreading pass for notation and grammar."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The integration is real: behavior-cloned initial grasps, a GMM stability estimator, and a self-attention policy that outputs continuous angle corrections. I haven't seen that exact combination in the prior work, so the novelty claim is fair, though it's an incremental integration rather than a new principle. The paper also deserves credit for doing real robot experiments with tactile sensors and showing that, on seven objects, the supported weight goes up when adaptation is on. The self-cited sensor hardware is not a red flag; it's independent engineering support.\n\nThe soft spots are exactly where the stress-test note lands. Table I is just seven rows of maxima. No trial counts, no error bars, no baseline that simply closes the gripper by a fixed Δθ when instability is detected. Since the action is an angle change, which changes grip force, a trivial force-increase reflex could reproduce the numbers. That's a load-bearing flaw, not a cosmetic one.\n\nThe other issue is the \"novel objects\" claim. The paper says 1200 demonstrations on 12 objects for training, then calls the Fig. 5 objects novel, but never lists the 12 training objects or states that the test set is disjoint. If any of the seven test objects appeared in training, the generalization case collapses. This needs to be stated explicitly.\n\nSmaller issues: Eq. 2 uses π_θ where Eq. 1 used π_w, and Eq. 6 is malformed. These are typos but should be fixed. The GMM threshold is selected from a ROC curve on training data, which is standard calibration, so I don't see circularity there.\n\nOverall, the architecture is coherent and the idea is sensible. The experimental evidence is underpowered for the claims made. The fixes are straightforward: report trial counts and variance, add a fixed-Δθ baseline, list the training objects, and release code and data. If those come through, this would be a solid contribution to tactile grasping.\n\nThis paper deserves a serious referee. It's not a desk reject, but it needs major revision before publication.","headline":"A plausible tactile-grasp adaptation pipeline, but the key generalization claim is undercut by missing experimental controls and training/test overlap ambiguity.","tokens_in":8228,"tokens_out":1461,"would_cite":false,"duration_ms":14956,"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":"The paper claims a demonstration-learned, tactile-only adaptive grasping policy can keep a two-finger gripper stable under pulling, vibration, and added weight, and that it raises the maximum holdable weight on all seven novel objects…","keywords":["tactile sensing","robust grasping","grasp adaptation","human demonstration","Gaussian mixture model","self-attention","parallel gripper"],"falsifier":"A decisive test would be to increase an object's weight remotely, for example by pumping liquid into a rigid container through a tube, so that mass changes while the tactile contact pattern and gripper angle stay unchanged; if the estimator reports stable and the policy issues no correction even though the grasp eventually slips, the premise that 32 tactile values plus gripper angle capture grasp stability is refuted.","tokens_in":7250,"feed_emoji":"🤖","tokens_out":8136,"duration_ms":67208,"temperature":0.7,"pith_summary":"The paper sets out to show that a robot gripper can respond to disturbances—a person pulling the object, vibration, or added weight—by adjusting its grip angle using only tactile readings, without vision or prior object knowledge. The adaptive policy is learned from human demonstrations of corrective gripping and is implemented as a stability estimator followed by an angle-correction network. On a two-finger electric gripper with 32-channel tactile fingertips, the authors test seven novel household objects and report that the adaptation increases the maximum weight the gripper can hold in every case. A sympathetic reading takes the central contribution to be evidence that continuous, tactile-only corrective action is a practical route to disturbance-resistant grasping.","feed_headline":"Tactile adaptation raises max stable grasp weight up to 79%","feed_subtitle":"Learned from human demos, the policy re-tightens gripper angle from tactile readings when objects slip, shake, or gain mass.","key_machinery":"The central machinery is a tactile grasp state $X^*=(S^*,\\theta^*,P^*)$, where $S^*\\in\\mathbb{R}^{32}$ is the pair of 16-channel fingertip readings, $\\theta^*$ is the gripper angle, and $P^*\\in\\mathbb{R}^7$ is the end-effector pose. A Gaussian mixture model over this state acts as a binary stability classifier: when the likelihood $p(X^*|\\Omega)$ falls below a threshold chosen from the ROC curve, the grasp is treated as unstable. A self-attention network—a mechanism that weights past time steps by relevance to the current output—then maps the sequence of tactile changes $(\\Delta S_i,\\Delta\\theta_i)$ to a corrective gripper-angle change $\\Delta\\theta$, updating $\\theta'\\mapsto\\theta+\\Delta\\theta$. Because the action space is just the gripper angle, the same learned policy can run iteratively on a parallel gripper until stability is restored.","core_discovery":"The paper's central claim is that continuous grasp adaptation can be learned entirely from human demonstrations and carried by tactile feedback: a two-finger gripper can detect an unstable grasp from the current tactile state and restore stability by adjusting only its finger angle. The trained policy, called he-AGPT, consists of a behavior-cloned initial grasp generator, a Gaussian-mixture stability estimator, and a self-attention adaptation network that outputs $\\Delta\\theta$. On seven novel objects—pillbox, tea can, mouthwash, milk bottle, wine bottle, perfume, and ink—the reported maximum holdable weight rises with adaptation in every case, for example from 34 g to 55 g for the pillbox and from 72 g to 129 g for the ink bottle. Live tests include a person pulling a milk bottle downward and water being poured into a wine bottle, with the fingers adjusting their angle during the disturbance.","pith_inferences":["A testable extension is to run the same demonstration pipeline on other two-finger grippers with different tactile layouts; if the angle-only action space transfers, the cost of deploying the policy on new hardware drops sharply.","The reported largest percentage gains on small rigid objects (pillbox and ink bottle) suggest that the value of angle correction may depend on object compliance; comparing correction magnitude against object stiffness would test that.","Because the policy uses no vision, it could be paired with a vision-based initial grasp planner while keeping tactile-only corrections, a combination that may address the abrupt-shape-change failures the paper itself notes.","An implicit implication is that human demonstrators encode a useful disturbance-response prior that could be mined for other tactile tasks such as in-hand manipulation, though the paper does not claim this."],"forward_implications":["For the seven tested objects, enabling the adaptation increases the maximum supported weight, with the largest relative gain on the ink bottle (72 g to 129 g).","The policy generalizes to novel objects of different sizes, shapes, and textures without retraining, using only tactile input and gripper angle.","Repeated corrective angle adjustments, rather than a single regrasp action, let the gripper ride out continuous disturbances such as pouring and pulling.","Because the adaptation loop needs no vision or object model, it can run on simple parallel grippers in settings where cameras are unavailable or occluded."],"supporting_citations":[{"why":"Supplies the closest prior baseline, an object-level impedance controller that uses tactile feedback to keep an Allegro hand stable under human pulling.","marker":"[14]"},{"why":"Documents the limitation of a single corrective action in tactile grasp correction, which the iterative angle adaptation here is designed to overcome.","marker":"[15]"},{"why":"Provides the grasp-template dictionary approach that the initial grasp generator adapts, replacing the visual sense with tactile feedback.","marker":"[18]"},{"why":"Supplies the self-attention mechanism used to turn sequences of tactile readings and angle changes into smooth corrective actions.","marker":"[19]"},{"why":"Describes the soft tactile palm with electrode array and conductive foam whose sensing principle underlies the fingertip tactile readings.","marker":"[20]"},{"why":"Describes the biomimetic tactile palm whose impedance-based contact sensing the fingertip sensors build on.","marker":"[21]"}],"fun_headline_variants":["Tactile policy increases stable grasp weight up to 79%","Gripper learns from human demos to adapt grip angle","Reactive tactile adjustment counters grasp disturbances","Adaptive grip from demos improves object hold","Touch-driven adaptation stabilizes robot grasp"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the 32 tactile readings and the gripper angle contain enough information to detect instability and to infer the corrective angle, so that no vision or object model is required.","fun_headline_variants_meta":{"raw":{"variants":["Tactile policy increases stable grasp weight up to 79%","Gripper learns from human demos to adapt grip angle","Reactive tactile adjustment counters grasp disturbances","Adaptive grip from demos improves object hold","Touch-driven adaptation stabilizes robot grasp"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000313,"raw_usage":{"total_tokens":1726,"prompt_tokens":839,"completion_tokens":887,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":455,"completion_tokens_details":{"reasoning_tokens":813}},"tokens_in":455,"tokens_out":887,"duration_ms":9335,"temperature":1.0,"reasoning_tokens":813,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T21:31:05.142044+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A decisive test would be to increase an object's weight remotely, for example by pumping liquid into a rigid container through a tube, so that mass changes while the tactile contact pattern and gripper angle stay unchanged; if the estimator reports stable and the policy issues no correction even though the grasp eventually slips, the premise that 32 tactile values plus gripper angle capture grasp stability is refuted.","supporting_citations":[{"cited_title":"Learning of grasp adaptation through experience and tactile sensing. in 2014 ieee,","cited_arxiv_id":null,"evidence_quote":"Supplies the closest prior baseline, an object-level impedance controller that uses tactile feedback to keep an Allegro hand stable under human pulling."},{"cited_title":"Estimating tactile data for adap- tive grasping of novel objects,","cited_arxiv_id":null,"evidence_quote":"Documents the limitation of a single corrective action in tactile grasp correction, which the iterative angle adaptation here is designed to overcome."},{"cited_title":"Learning a dictionary of prototypical grasp-predicting parts from grasping experience,","cited_arxiv_id":null,"evidence_quote":"Provides the grasp-template dictionary approach that the initial grasp generator adapts, replacing the visual sense with tactile feedback."},{"cited_title":"A novel tactile palm for robotic object manipulation,","cited_arxiv_id":null,"evidence_quote":"Describes the soft tactile palm with electrode array and conductive foam whose sensing principle underlies the fingertip tactile readings."},{"cited_title":"A biomimetic tactile palm for robotic object manipulation,","cited_arxiv_id":null,"evidence_quote":"Describes the biomimetic tactile palm whose impedance-based contact sensing the fingertip sensors build on."}],"review_version":1}