{"id":"9cbee699-5af6-4aad-b717-d2674045b002","arxiv_id":"2608.07600","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"AdaDexGrasp learns to fuse point clouds with finger-level tactile labels to generate, judge, and correct dexterous grasps, reporting 91%/82%/83% success on seen, unseen-object, and unseen-category sets in simulation.","lead":"A robot hand system uses both 3D vision and fingertip touch to plan a first grasp, then adjusts the grip after contact if the touch signals look unstable. In simulation and on a real robot arm, it reports higher grasp success than ten comparison methods, especially on objects never seen in training.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Sim-to-real hand mismatch: Table 3's real-world results cannot be attributed to the described method without a disclosed 22-to-11 DoF mapping or retraining protocol.","rationale":"The reader's weakest assumption is exactly the gap between the 22-DoF simulated ShadowHand and the 11-DoF real Psibot SynHand. My stress-test confirms this is the most load-bearing concern. The paper's own text provides only a single vague sentence in Section 3.3 about adapting to different hands, with no concrete mechanism. The supplementary material's discussion of left- versus right-arm deployment does not close the gap because arm mirroring and joint-space dimensionality are different issues. The real-world results in Table 3 are central to the claim of practical, generalizable grasping; without a documented retraining or mapping step, those numbers are unverifiable and the method as written cannot be reproduced on the stated hardware. This does not make the paper fraudulent or internally contradictory in a way that forces rejection; the gap is addressable by adding the missing protocol. Therefore the reader's CONDITIONAL verdict remains appropriate: accept only if the authors disclose the exact SynHand adaptation procedure and ideally release code or checkpoints. My proposed test is a direct check of which model produced Table 3, which settles whether the concern is a documentation gap or a genuine failure of the described method.","tokens_in":13951,"tokens_out":5808,"duration_ms":60675,"concrete_test":"Inspect the model checkpoint used for Table 3 and check the output dimension of the pose generator and adaptation model. If the checkpoint uses a 22-dimensional joint vector, attempt to run it directly on the 11-DoF SynHand without modification; if it cannot be loaded or success collapses on the 260 real trials, the real-world result is not produced by the described method. If the checkpoint uses an 11-dimensional joint vector, require the authors to disclose the SynHand training pipeline, simulator, data collection, and hyperparameters, and confirm that the method sections describing j in R^22 and Eq. (8) apply to that separately trained model.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central real-world claim rests on an unstated embodiment transfer. Section 3.1 defines the hand state as s = [p, q, j] with j in R^22, and the adaptation pairing distance in Eq. (8) uses that 22-dimensional joint vector. The real platform in Section 4.4 and Figure 6 is an 11-DoF Psibot SynHand-6. The only transition sentence is in Section 3.3: training data 'can be adapted to different dexterous hands by training the policy on the corresponding hand models.' No retraining protocol, no architecture change for the output dimension, no SynHand simulation environment, and no kinematic mapping from 22 to 11 joints is described anywhere in the paper or supplement. The supplementary's note that left-arm and right-arm results validate hardware agnosticism does not address the 11-vs-22 DoF mismatch; mirroring a kinematic chain is not the same as changing its joint space. Consequently, Table 3's 90/87/81 success rates could only have been produced either by a separate SynHand-trained model that is never documented, or by an unspecified mapping that is never defined. Either way, the reported real-world results are not connected to the method as specified. This is load-bearing because the abstract and conclusion claim real-world validity and generalization, and the real-world table is the primary evidence for transfer.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AdaDexGrasp, a dexterous grasping framework that fuses visual point clouds with tactile feedback. The system comprises three learned components: a contact-driven grasp pose generator that predicts a per-point contact map carrying finger/palm identities (Section 3.3); a grasp feasibility classifier that operates on a tactile-mapped point cloud (Section 3.4); and a diffusion-based pose adaptation model that refines suboptimal poses (Section 3.5). These components are combined into a closed-loop refinement procedure (Section 3.6, Eqs. 12-14) that iterates until the classifier predicts success or a maximum iteration count is reached. Training data are collected from PPO rollouts in an IsaacGym environment with a 22-DoF ShadowHand (Section 3.7). Experiments report 91%/82%/83% success on seen, unseen-object, and unseen-category sets in simulation (Table 1) and 90%/87%/81% on a real 11-DoF Psibot SynHand-6 across 26 objects (Table 3), with ablations in Table 2 and additional quantitative analysis in the supplementary material.","tokens_in":14242,"tokens_out":8279,"duration_ms":70222,"significance":"If the reported results are reproducible, the framework advances the state of the art on an important and under-addressed problem: tactile-guided closed-loop dexterous grasping. Strengths include the integration of grasp generation, feasibility prediction, and adaptive refinement in one pipeline; a semantically labeled contact map that ties tactile signals to finger identities; consistent ablations showing that both contact IDs and adaptation matter (Table 2); a new IsaacGym tactile environment; and real-world experiments with a disclosed protocol (260 trials per method, supplementary Section E). The supplementary also provides honest quantitative ablations of the intermediate contact map, the pairing metric, and the tactile representation choice. The empirical claims are, however, limited by the issues in the major comments, in particular the unexplained 22-to-11 degree-of-freedom transfer between the simulated hand and the real hand, which disconnects Table 3 from the method as specified, and the absence of variance or significance reporting.","major_comments":[{"comment":"The real-world results in Table 3 cannot be attributed to the method as specified. Section 3.1 defines the hand state j as a 22-dimensional joint vector, Eq. (8) computes the adaptation pairing distance on that 22-dimensional joint space, and all components are trained with the 22-DoF ShadowHand described in Section 4.1. The real platform in Section 4.4 and Figure 6 is an 11-DoF Psibot SynHand-6. The only cross-embodiment statement, Section 3.3 ('training data ... can be adapted to different dexterous hands by training the policy on the corresponding hand models'), does not describe a kinematic mapping from 22 to 11 joints, a retraining protocol for the SynHand, an architecture change for the output dimension of Gpose or the adapter, or a SynHand simulation environment. The supplementary's left-arm/right-arm validation concerns mirror symmetry of a kinematic chain and does not address joint-space dimensionality. The authors must disclose how the Table 3 numbers were produced: either a separately trained SynHand model (whose training and simulation setup must then be documented) or an explicit joint mapping (which requires revising Eq. (8) and the pose-output layers, and re-verifying the pairing and adaptation losses). Without this disclosure, the real-world success rates do not support the central claim of real-world validity and generalization.","section":"§3.1, §3.5, §4.4, Eq. (8)"},{"comment":"All success rates are reported as single point estimates without error bars, number of simulation trials or seeds, or significance tests. The supplementary (Section E) discloses 260 real-world trials per method, but no analogous count is given for simulation, and none of the tables report variance. This matters for the paper's headline claim ('significantly enhances grasp success rates', abstract) because several decisive margins in Table 2 are small (e.g., 84% vs 91% on seen objects for 'w/o contact id in adaptation'; 72% vs 74% on unseen objects between 'Object-only' and 'w/o PC'). The authors should report per-split trial counts, at least three seeds for the main simulation comparisons, confidence intervals, and a significance test for the principal comparisons in Tables 1 and 2.","section":"Tables 1–3; abstract"},{"comment":"The simulation asset description is internally inconsistent. The text states 'including 50 objects from 6 categories. Among them, 20 objects are used for training and 30 unseen-category objects for testing', which implies the 50 objects are partitioned into 20 training and 30 unseen-category test objects, leaving no test objects from seen categories. Yet Tables 1 and 2 report a separate 'Unseen Objects' column defined in Section 4.1 as 'objects from the same categories but unseen in training'. As written, the 'Unseen Objects' column has no defined test set. The authors should specify the exact number of objects in each of the three splits and how the 6 categories are partitioned.","section":"§4.1"},{"comment":"The closed-loop stopping rule in Eq. (13) uses the classifier's own prediction with threshold τ_succ = 0.5, and the adapter's loss in Eq. (11) contains the term λ(1 − f_θ(...)), which encourages refined poses to score highly on that same classifier. The paper does not report whether classifier confidence is calibrated against physical stability (the physics-based success metric defined in Section 4.1), nor whether the classifier is retrained after the adapter is trained. If the classifier is biased, the adapter can exploit that bias, and the supplementary's statement that 'over 90% converge to a successful grasp within fewer than 3 refinement iterations' is ambiguous as to whether convergence is judged by the classifier or by physical lift stability. Please report classifier-versus-physics agreement on the simulation test sets and clarify the arbiter behind the convergence statistic.","section":"§3.5–§3.6, Eq. (11), supplementary Section B"}],"minor_comments":[{"comment":"The claim 'gains also persist at scale: ... our method still reaches 87%' on the DexGraspAnything dataset lacks a baseline number and a split definition; please specify what 87% refers to and compare against a baseline at the same scale.","section":"§4.2"},{"comment":"The sentence 'This estimates 0 serves as an initial grasp pose' contains a typo ('This estimates 0' should be 'This estimate s(0)' or similar).","section":"§3.3"},{"comment":"The success criterion differs between simulation (stable for 1 second) and real world (3 seconds); please state the rationale so readers can gauge comparability.","section":"§4.1 vs §4.4"},{"comment":"'Avg. TIF' is used without definition; the text should define the averaged tactile intensity feature and its normalization.","section":"Fig. 7"},{"comment":"Several baselines are named without citations or precise configuration details (e.g., Robot Synesthesia, Intuitive Closed-Loop, DexGraspVLA); please add references or definitions so the comparisons are reproducible.","section":"§4.2 baselines"}],"recommendation":"major_revision","confidential_remarks":"The DoF mismatch between the simulated 22-DoF ShadowHand and the real 11-DoF SynHand-6 is the main barrier between this manuscript and a publishable claim; I recommend the editor treat the requested disclosure as mandatory. The paper otherwise appears to be a competent systems contribution with honest ablation content. I have no concerns about citation practice beyond the usual arXiv-dense reference list."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read the AdaDexGrasp preprint. The core idea is worth your time: they attach finger/palm identities to a predicted contact map for initial grasp generation, then close the loop with a tactile-mapped point cloud that a classifier and a diffusion-based adapter use to refine the pose after contact. That specific combination is new—the baselines they cite don't do it—and the sim results are clean. On seen/unseen-object/unseen-category splits they get 91/82/83%, beating the strongest baseline (DexGraspAnything, 77/72/67) by double digits, and the ablation table shows consistent drops when you remove contact IDs or the adaptation loop. The gains grow on unseen categories, which is the right place for a generalization claim to show.\n\nThe load-bearing problem is the sim-to-real hand. The whole method is defined on j in R^22, including the pairing distance in Eq. (8). The real robot is an 11-DoF Psibot SynHand. The only bridge is one sentence in Sec. 3.3 about training on different hand models—no retraining protocol, no architecture change for the output dimension, no SynHand simulation, no kinematic mapping. The supplementary's left-arm/right-arm test does not address a change in DoF. So the real-world table (90/87/81) cannot be attributed to the method as described. Either there's a separate SynHand-trained model that's never documented, or there's an undefined mapping. Either way, the real-world evidence, which is central to the abstract's claim, is not connected to the system.\n\nSecondary but addressable: no error bars or significance tests anywhere; the real-world protocol reports 260 trials per method but no variance; one baseline (RobotSynesthesia) has no citation or description; no code or data. These are fixable.\n\nWho should read this? Anyone working on dexterous grasping or visuo-tactile fusion. The simulation framework is solid and the adaptive loop is a useful design. But I would not take the real-world numbers at face value until the hand mapping is explained. I'd send it to peer review with a major-revision expectation—the core idea deserves scrutiny, and the sim results are a legitimate data point. If the authors can produce the SynHand retraining details or rule out the mismatch, the central claim could hold.","headline":"Interesting simulation results and a clean ablation, but the real-world numbers are not connected to the described method because of an unexplained 22-to-11 DoF hand mismatch.","tokens_in":14787,"tokens_out":5055,"would_cite":false,"duration_ms":44551,"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 visuo-tactile representation that binds touch to finger identities lifts dexterous grasp success to 91% in simulation and 90% on a real robot.","keywords":["dexterous grasping","visuo-tactile fusion","tactile feedback","contact map","grasp adaptation","closed-loop refinement","point cloud","sim-to-real transfer"],"falsifier":"Run the same closed-loop policy on the real 11-joint hand after simply ignoring or remapping the 11 extra simulated joints; if grasp success then falls to the level of the vision-only baselines, the claimed sim-to-real transfer of the tactile-guided framework is refuted.","tokens_in":13769,"feed_emoji":"🖐️","tokens_out":8525,"duration_ms":71219,"temperature":0.7,"pith_summary":"This paper tries to establish that dexterous grasping improves sharply when tactile feedback is not just an extra input channel but is spatially bound to specific fingers and the palm. The proposed framework generates an initial grasp from a point cloud by predicting which hand part should touch which region of the object, then uses measured contact to classify whether the grasp is stable and, if not, to iteratively correct the hand pose. In simulation the system reaches 91% success on seen objects, 82% on unseen objects, and 83% on unseen categories; on a real robot with a multi-finger hand it reports 90%, 87%, and 81%. If correct, this indicates that post-contact touch information, not better visual planning alone, is the key to generalizing to novel object shapes. This matters because nearly all current grasping systems are open-loop: they commit to a pose from vision and never correct once the fingers touch the object.","feed_headline":"Tactile-guided grasping hits 91%, beating vision-only","feed_subtitle":"A closed-loop controller reads fingertip contact and corrects loose grasps, generalizing to unseen objects.","key_machinery":"The central mechanism is the tactile-mapped point cloud $\\mathcal{P}_{vt}$, which is the object point cloud annotated per point with RGB color, a scalar contact state $k_i$, a six-channel tactile intensity vector $c_i$, and a categorical contact ID that identifies the specific finger or palm in contact. This fused representation is the unified input to all three models: a contact-driven generator that first predicts a semantic contact map $M$ and then a grasp pose, a grasp success classifier $f_\\theta$, and a closed-loop adaptation model $g_\\phi$ that produces corrective pose updates. The contact ID is what binds vision to touch: it tells the network not just that contact happened but which hand part made it, which enables fine-grained finger-object reasoning during both planning and refinement.","core_discovery":"The paper's central claim is that a unified visuo-tactile representation, in which tactile signals are bound to finger and palm identities, enables a single pipeline to generate contact-aware initial grasps, predict whether they will succeed, and iteratively correct unsuccessful ones after contact. The reported results are 91% success on seen objects, 82% on unseen objects, and 83% on unseen categories in simulation, and 90%, 87%, and 81% respectively on a real robot. The authors attribute the gains to the semantic binding of contact to hand parts: removing that binding in generation drops seen-object success from 91% to 79%, and disabling the adaptation module drops unseen-category success from 83% to 59%. This leads them to conclude that post-contact tactile feedback, not just better visual planning, is the decisive factor for generalizable dexterous grasping.","pith_inferences":["A testable extension is to apply the same contact-ID binding to a two-finger gripper, since the mechanism does not depend on five-finger dexterity.","The paper's own failure analysis points to slippage and size-limit cases; adding explicit slip detection or pre-grasp size estimation could be a direct next step.","The 22-joint training state versus the 11-joint real hand leaves the sim-to-real transfer under-specified; I would expect retraining on the real hand's kinematics to change the reported real-world numbers."],"forward_implications":["Enabling the adaptation module raises unseen-category success in simulation from 59% to 83%, so tactile-guided correction is the main source of generalization.","Removing finger/palm identity from the contact map during generation drops seen-object success from 91% to 79%; the semantic binding is load-bearing.","On a dataset with more than 15,000 objects, the framework retains 87% success, suggesting the representation scales beyond the 50-object training set.","In practice, refinement is triggered in about 30% of trials and over 90% of those converge within three iterations, so the closed loop is fast enough for real-time use.","Real-world results track simulation trends, supporting the view that the same visuo-tactile representation works outside the simulator."],"supporting_citations":[{"why":"Provides the GPU-accelerated physics simulator used for all simulated training and evaluation.","marker":"[23]"},{"why":"Supplies the reinforcement-learning data collection setup that produces paired successful and failed grasp states used to train all components.","marker":"[43]"},{"why":"Supplies the proximal-policy-optimization algorithm used to generate the training trajectories.","marker":"[31]"},{"why":"Provides the object asset set and serves as a purely visual baseline that the framework must beat.","marker":"[33]"},{"why":"Provides the 15k-object dataset used for the large-scale generalization test and serves as a strong vision-only baseline.","marker":"[54]"},{"why":"A diffusion-based baseline that generates grasps from object point clouds; it is compared against to show the value of tactile refinement.","marker":"[39]"},{"why":"A baseline that also predicts contact maps; comparison highlights the benefit of adding finger-identity semantics to contact maps.","marker":"[51]"}],"fun_headline_variants":["AdaDexGrasp: vision-touch fusion boosts grasp success to 91%","Touch-corrected dexterous grasping: 91% success, adapts to objects","Feel your grip: tactile feedback boosts dexterous grasp to 91%","Vision plus touch: adaptive dexterous grasping hits 91%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the 22-dimensional joint state and contact patterns learned in simulation transfer to the real 11-joint hand, since the paper defines the state in 22 joints but evaluates on an 11-joint hand without describing a kinematic alignment or retraining step.","fun_headline_variants_meta":{"raw":{"variants":["AdaDexGrasp: vision-touch fusion boosts grasp success to 91%","Touch-corrected dexterous grasping: 91% success, adapts to objects","Feel your grip: tactile feedback boosts dexterous grasp to 91%","Vision plus touch: adaptive dexterous grasping hits 91%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000653,"raw_usage":{"total_tokens":2957,"prompt_tokens":873,"completion_tokens":2084,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":489,"completion_tokens_details":{"reasoning_tokens":2006}},"tokens_in":489,"tokens_out":2084,"duration_ms":15531,"temperature":1.0,"reasoning_tokens":2006,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:30:07.632193+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same closed-loop policy on the real 11-joint hand after simply ignoring or remapping the 11 extra simulated joints; if grasp success then falls to the level of the vision-only baselines, the claimed sim-to-real transfer of the tactile-guided framework is refuted.","supporting_citations":[{"cited_title":"08257 20 X","cited_arxiv_id":null,"evidence_quote":"Provides the 15k-object dataset used for the large-scale generalization test and serves as a strong vision-only baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"A baseline that also predicts contact maps; comparison highlights the benefit of adding finger-identity semantics to contact maps."}],"review_version":1}