{"id":"59323abc-2264-4804-a580-8dec161e852f","arxiv_id":"2506.14317","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A simulation-trained teacher-student policy achieves zero-shot sim-to-real closed-loop target-oriented dexterous grasping in cluttered scenes, with 83.9 percent real-world success.","lead":"This paper presents a robot system that can pick a specific target object out of a cluttered pile using a dexterous hand, trained entirely in simulation and deployed in the real world without real-world demonstration data. The reported real-world success rate is 83.9 percent over 167 attempts, which makes this an interesting step toward practical dexterous manipulation in unstructured scenes.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Real-world success metric in Appendix E.1 may inflate the 83.9% rate: success is declared by a finger-torque threshold, not by verified lift, so the headline zero-shot claim rests on an unvalidated criterion.","rationale":"The reader identified the finger-torque success signal as the weakest assumption, and I agree that this is the most load-bearing concern. The paper's headline real-world result is 83.9% success over 167 attempts, and the only described mechanism for determining success in the real world is the torque threshold in Appendix E.1. The simulation success definition is explicit ('the target object is lifted 0.1 meters'), while the real-world definition appears to declare success before the lift-up action is verified. This asymmetry means the reported real-world number may not measure the same behavior that is validated in simulation, and it is exactly the kind of hidden evaluation detail that can systematically inflate a success rate. The concern is not that the authors are dishonest; it is that the metric is under-specified and the missing threshold value is a tunable parameter. A focused re-analysis of the videos with an external lift-based criterion would settle the issue. Because the rest of the paper has strong simulation ablations and a plausible sim-to-real pipeline, I would not reject the work outright, but the central zero-shot claim should be conditioned on validation of the real-world metric.","tokens_in":15687,"tokens_out":3821,"duration_ms":41872,"concrete_test":"Re-label all 167 real-world attempts from the released videos (or a re-run) using an external criterion: the object centroid must remain above 0.1 m for at least 1 s after the predefined lift-up action, or a human labeler must confirm the object is fully lifted and held. Compute the corrected success rate and report the torque threshold used. Additionally, apply the same torque-threshold success rule to the simulation evaluation and compare it with the lift-0.1 m metric; if the torque rule overestimates simulation success by more than 5 percentage points, the real-world number is likely inflated.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Appendix E.1 states that 'we use finger torque as a success signal: Once it exceeds a threshold, the grasp is considered successful and a predefined lift-up action is executed.' The simulation metric requires the target to be lifted 0.1 m; the real-world metric appears to count success at the torque event before the lift-up is completed or verified. If the torque threshold is reached during a momentary or partial contact, the trial is permanently labeled a success even if the object is not lifted or is dropped during the predefined lift. The threshold value is not reported, and no independent or human verification of the 167 attempts is described. Because the central claim is that the system achieves 83.9% real-world success and robust performance in clutter, an inflated success metric would directly undermine the empirical support for zero-shot transfer. The concern is concrete: the evaluation criterion is not aligned with the simulation success definition, and the missing threshold value is a free parameter that could have been tuned to raise the reported rate.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes ClutterDexGrasp, a two-stage teacher-student framework for target-oriented dexterous grasping in cluttered tabletop scenes. A privileged teacher policy is trained with PPO using a geometry-and-spatial distance representation, a clutter-density curriculum, and an interaction-safety curriculum; a student 3D diffusion policy (DP3) is then distilled offline from teacher rollouts on partial point-cloud observations. Simulation evaluation covers 88 training objects, 2,029 unseen test objects, three clutter densities, and three seeds, reporting success rates around 87--92% for the teacher and 74--91% for the student, with ablations of the representation, curriculum, and safety components. Real-world evaluation reports 83.9% success over 167 attempts on 41 unseen objects without real-world demonstrations, and the paper claims this is the first zero-shot sim-to-real closed-loop system for target-oriented dexterous grasping in cluttered scenes.","tokens_in":15975,"tokens_out":3617,"duration_ms":37959,"significance":"If the central claims hold, this is a strong and practically relevant systems contribution: it demonstrates that an RL-trained teacher and a point-cloud diffusion student can transfer to a real 7-DoF arm and 12-DoF hand in cluttered scenes without real-world demonstrations. The simulation results are carefully reported with three seeds, a large unseen-object evaluation, and informative ablations; the appendices give substantial implementation detail, and the authors provide a project website with videos. The main weakness is the real-world evidence: the success criterion is an internal finger-torque threshold rather than an independently verified lift, no real-world baselines or error bars are reported, and the \"zero-shot\" label is weakened by system identification and several real-world-specific adjustments. These issues bear directly on the headline 83.9% success rate and on the novelty claim of zero-shot transfer, so the verdict is major revision rather than acceptance.","major_comments":[{"comment":"The text says the student \"shows less than a 5% average success rate drop across all combinations of seen and unseen objects and layouts,\" but the table shows a 7.4 percentage-point drop for seen objects with unseen layouts (80.9 to 73.5) and 7.4 for unseen objects with unseen ultra-dense layouts (81.6 to 74.2). The average may be below 5%, but the claim as written is misleading for individual cells. Please state the per-cell drops explicitly or rephrase to \"average drop\" with the individual maximum drop disclosed.","section":"Table 1, Section 5.1.1"}],"minor_comments":[{"comment":"References [22] and [23] are the same paper (Ross, Gordon, Bagnoli, \"A reduction of imitation learning and structured prediction to no-regret online learning\") and are cited with identical text; please merge them or remove the duplicate.","section":"Appendix A/References"},{"comment":"The framework diagram is dense and some text is very small (e.g., \"Force Penalty2\" and \"Table Contact Penalty1\"); consider enlarging or separating the three curriculum stages into distinct panels for readability.","section":"Figure 2"},{"comment":"The action space description says the hand has 6-DoF actuated finger joints and 6-DoF underactuated joints, but the action space is defined as 7D arm joint changes plus 6D absolute hand joint positions. Please clarify which joints are actually commanded and how the underactuated joints are modeled in both simulation and real hardware.","section":"Section 3"},{"comment":"The point cloud observation is described as O_pc in R^{4x3584} plus R^{4x512} and R^{4x1024}, totaling 5120 points, but the student observation dimension is given as R^{4109} in Appendix C. Please reconcile the dimension (joints 13 + 4096 points + masks?).","section":"Appendix C.1"},{"comment":"The limitation paragraph mentions tiny objects but does not mention the failure mode identified in Appendix F where objects fall outside the arm's working range; including that in the main text would strengthen the honesty of the failure analysis.","section":"Section 7"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid systems contribution with careful simulation work, but the real-world evaluation needs strengthening before the central zero-shot claim can be fully credited. The torque-threshold success criterion is the most serious issue; if it cannot be validated retroactively, the authors should either re-run the real-world evaluation with an external success criterion or substantially soften the claims. I would not reject the paper outright because the simulation evidence and the engineering pipeline are valuable and the issues appear addressable within a major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid system paper with a genuinely new integration of clutter-aware representation, density curriculum, safety curriculum, and DP3 distillation. The simulation work is the strongest part: three seeds, large unseen-object evaluation (2029 objects), and clean ablations showing the density curriculum is necessary (from-scratch clutter training gets 0%) and the safety curriculum cuts contact force by roughly half with only a 1.9% success drop. That is real evidence.\n\nThe real-world evaluation is where I part ways with the authors' confidence. Appendix E.1 says success is declared when finger torque exceeds a threshold, before the lift-up is verified. The simulation success criterion is lifting the target 0.1 m. So the headline 83.9% over 167 attempts rests on a metric that may count momentary or partial contacts as successes. The threshold value is not reported, and the authors note that they added this signal because \"the system occasionally failed to detect successful grasps.\" That makes the concern concrete, not hypothetical. Also, \"zero-shot\" is diluted by system identification and by slowing down the end-effector in the real world. And there are no real-world baselines or error bars. None of this invalidates the simulation results, but it weakens the paper's central real-world claim.\n\nThe stress-test note about the torque metric holds up. I'd want the authors to align the real-world success detection with the sim lift criterion, for example by verifying the lift with a second signal, reporting the threshold, and giving per-scene breakdowns. If they can do that, the paper becomes substantially stronger.\n\nWho is this for? People working on dexterous sim-to-real and teacher-student grasping. It deserves serious peer review, not desk rejection. I'd recommend accept-with-major-revisions over the current submission, mainly on the strength of the sim evidence and the honest failure analysis in Appendix F. The metric issue is fixable, and the system-level contribution is real.","headline":"Solid sim-to-real system with strong simulation ablations, but the real-world success metric (torque threshold before lift verification) likely inflates the headline 83.9% and needs fixing before publication.","tokens_in":16477,"tokens_out":2259,"would_cite":true,"duration_ms":27031,"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 that a policy trained entirely in simulation can close the loop on target-oriented dexterous grasping in cluttered real-world scenes, with no real-world demonstrations.","keywords":["Dexterous grasping","Cluttered scenes","Sim-to-real transfer","Teacher-student distillation","3D diffusion policy","Curriculum learning","Point cloud policy","Closed-loop manipulation"],"falsifier":"Run the same real-world scenes with an independent success criterion, for example a vision system or human judge confirming that the target object remains in the hand and is lifted 0.1 meters after the robot stops, without contact support from the table or other objects, and compare the resulting success rate with the torque-threshold-based 83.9 percent. A large gap would show the reported metric overcounts.","tokens_in":15510,"feed_emoji":"🦾","tokens_out":5408,"duration_ms":52461,"temperature":0.7,"pith_summary":"This paper tries to establish that a dexterous hand-arm system can learn to grasp a specified target object in dense clutter entirely in simulation, then transfer to the real world with no real-world demonstrations. The proposed ClutterDexGrasp trains a teacher policy with privileged geometric information and staged curricula, distills it into a point-cloud-based 3D diffusion policy, and reports an 83.9 percent success rate over 167 real-world attempts. The relevance is that closed-loop, target-oriented dexterous grasping in occlusion-heavy scenes has previously required either open-loop pose prediction or large amounts of real human demonstrations; this work claims a path that avoids both. It also reports that the student policy loses under five percentage points of success relative to the teacher in simulation, including on unseen objects and ultra-dense layouts.","feed_headline":"Sim-trained policy grasps real clutter targets 83.9% of the time","feed_subtitle":"Teacher-student pipeline transfers dexterous hand-arm control to unseen objects and layouts with zero real-world demos.","key_machinery":"The load-bearing mechanism is the teacher-student distillation pipeline. The teacher is an RL policy trained with PPO on a geometry-and-spatial representation: for each of 11 selected finger links, the 3D distance vector to the nearest sampled points on the target object ($d_{\\text{pos}}$) and on non-target objects ($d_{\\text{neg}}$). These feed a reward that encourages approach to the target and penalizes risky closeness to clutter. Three curricula shape the behavior: single-object grasping first, then cluttered-scene fine-tuning, then a safety stage that tightens a fingertip force threshold and terminates episodes on high force or table penetration. The student is a 3D diffusion policy (DP3) that acts on a 4096-point partial point cloud with a target mask, augmented with synthetic robot and ground point clouds, and is trained offline on 20,000 teacher rollouts. System identification calibrates arm and hand dynamics so that the student transfers without real-world fine-tuning.","core_discovery":"The central claim is that the combination of a privileged geometry-and-spatial representation, a clutter-density curriculum, and an interaction-safety curriculum makes reinforcement learning for dexterous cluttered-scene grasping stable enough to be distilled into a real-world-ready student. The teacher observes, for each finger link, the 3D distance to the nearest sampled surface points of the target object and of non-target objects; this compact representation lets the policy learn collision-aware, human-like strategies such as gently nudging occluders aside before grasping. The student replaces privileged state with a partial point cloud and is trained by imitation on 20,000 successful teacher trajectories, using a 3D diffusion policy. The paper claims this is the first zero-shot sim-to-real closed-loop system for target-oriented dexterous grasping in cluttered scenes, with real-world success at 83.9 percent over 167 attempts and robust generalization to unseen objects and layouts.","pith_inferences":["We infer that the same teacher-student pattern, privileged distance representation plus staged curricula plus diffusion distillation, could transfer to other contact-rich dexterous tasks such as in-hand reorientation or tool use, provided a comparable privileged signal exists in simulation.","The reported success metric depends on a finger-torque threshold; an independent test in which a human or vision system confirms the object stays lifted after the robot raises the arm would settle whether the 83.9 percent figure reflects stable grasps.","The failure analysis points to a concrete extension: multi-view or active perception, rather than a single side camera, should reduce failures on tiny or heavily occluded objects without changing the policy-learning pipeline.","The data-scaling experiment suggests the student's generalization improves with trajectory count, implying the method's ceiling is partly a data-generation question, not only an algorithmic one."],"forward_implications":["If the central claim holds, real-world dexterous grasping in clutter no longer requires collecting human demonstrations for every new scene; a simulation-trained policy can be deployed directly.","The under-5% teacher-student success drop suggests that point-cloud distillation preserves the strategic behaviors, such as occluder clearing and side approaches, learned by the privileged teacher.","Because the policy generalizes to ultra-dense scenes it never saw in training, the clutter-density curriculum appears to teach a scalable skill rather than a memorized layout.","The explicit safety curriculum implies that sim-to-real dexterous policies can be made gentle enough for contact-rich interaction, not just high-scoring in simulation.","A practical consequence is that target selection can be interactive: the system uses SAM2 to mask the target object in the camera view, so a user can choose which object to grasp."],"supporting_citations":[{"why":"Supplies the distance-between-hand-links-and-object representation that the teacher observation extends to cluttered scenes.","marker":"[26]"},{"why":"Demonstrates the teacher-student paradigm for single-object dexterous grasping that this paper adapts to clutter.","marker":"[27]"},{"why":"Provides the DP3 3D diffusion policy backbone used as the student policy.","marker":"[35]"},{"why":"Provides the diffusion policy framework for visuomotor action generation used in the student.","marker":"[49]"},{"why":"Introduces synthetic robot point-cloud augmentation used to bridge the perception gap.","marker":"[7]"},{"why":"Supplies the system identification procedure used to calibrate arm and hand dynamics for sim-to-real transfer.","marker":"[60]"},{"why":"Provides the training object set of 88 objects from GraspNet1Billion used to build training scenes.","marker":"[61]"},{"why":"Provides the Omni6DPose object set used to evaluate generalization to unseen objects.","marker":"[62]"},{"why":"Supplies SAM2, used to segment the target object in real-world point clouds for interactive target selection.","marker":"[64]"}],"fun_headline_variants":["Zero-shot sim-to-real dexterous grasping hits 83.9% in clutter","83.9% real-world grasp success via sim-only training","First zero-shot closed-loop dexterous grasping in clutter","Dexterous hand-arm policy transfers sim to real without demos"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the real-world success signal, a finger-torque threshold that triggers a lift-up, correctly identifies genuine, stable grasps; if that threshold counts partial or unstable contacts as successes, the reported 83.9 percent rate would overstate the policy's real ability.","fun_headline_variants_meta":{"raw":{"variants":["Zero-shot sim-to-real dexterous grasping hits 83.9% in clutter","83.9% real-world grasp success via sim-only training","First zero-shot closed-loop dexterous grasping in clutter","Dexterous hand-arm policy transfers sim to real without demos"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000966,"raw_usage":{"total_tokens":4127,"prompt_tokens":981,"completion_tokens":3146,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":597,"completion_tokens_details":{"reasoning_tokens":3069}},"tokens_in":597,"tokens_out":3146,"duration_ms":22055,"temperature":1.0,"reasoning_tokens":3069,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:18:00.483099+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same real-world scenes with an independent success criterion, for example a vision system or human judge confirming that the target object remains in the hand and is lifted 0.1 meters after the robot stops, without contact support from the table or other objects, and compare the resulting success rate with the torque-threshold-based 83.9 percent. A large gap would show the reported metric overcounts.","supporting_citations":[{"cited_title":"Zhang, S","cited_arxiv_id":null,"evidence_quote":"Supplies the distance-between-hand-links-and-object representation that the teacher observation extends to cluttered scenes."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the DP3 3D diffusion policy backbone used as the student policy."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces synthetic robot point-cloud augmentation used to bridge the perception gap."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the system identification procedure used to calibrate arm and hand dynamics for sim-to-real transfer."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the training object set of 88 objects from GraspNet1Billion used to build training scenes."},{"cited_title":"Zhang, W","cited_arxiv_id":null,"evidence_quote":"Provides the Omni6DPose object set used to evaluate generalization to unseen objects."}],"review_version":1}