{"id":"820d54a0-cc57-429d-98b3-0886b8b980da","arxiv_id":"2508.00491","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An imitation-learning diffusion policy controlling wrist and finger movements from an eye-in-hand camera grasps diverse objects with the Hannes prosthetic hand in three real-world scenarios.","lead":"A team at the Italian Institute of Technology trained a diffusion policy on 450 demonstrations to control the wrist and fingers of the Hannes prosthetic hand from an in-hand camera, achieving about 79% grasp success on known objects and 76% on new objects. The work offers a path to reducing the cognitive load of prosthetic users by automating grasp execution from visual input.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Success is measured as a human-plus-policy task, so the claim that the learned policy is responsible for wrist/hand control is not yet isolated; a scripted no-policy baseline is needed.","rationale":"The paper's core feasibility claim is plausible and the hardware evaluation is substantial, but the central assertion rests on success rates from a task in which the human user and the policy act together. The reader identified the human-in-the-loop protocol as the weakest point but framed it mainly as an amputee-transfer question; I see a more immediate attribution problem: without a no-policy or policy-perturbation control, the success rates cannot be assigned to the learned policy. This does not imply any impropriety; it is a standard missing-control concern. I also note the Sec IV-D comparison with the visual-servo baseline uses different success criteria, so it should not be used to infer policy skill. A scripted baseline plus per-object significance testing would settle whether the diffusion policy is doing the load-bearing work. The reader's CONDITIONAL verdict remains appropriate, with the explicit condition that such an attribution control be added; the amputee-transfer question is a further, separate limitation that future user studies should address.","tokens_in":11984,"tokens_out":9376,"duration_ms":107430,"concrete_test":"Run a matched no-policy control in all three scenarios using the same operators, objects, and trial counts: have the user perform the same arm approach while (a) the wrist is held at a fixed neutral configuration and (b) hand closure is triggered by a simple time/proximity rule instead of the diffusion policy. If this scripted baseline achieves success within statistical noise of HannesImitationPolicy's 79.3%/76%, the policy-specific contribution is not demonstrated; if it fails substantially on the same objects, the learned wrist/hand actions are load-bearing and the original attribution stands. Reporting per-object success and a significance test (e.g., exact binomial or permutation test) would make the comparison decisive.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"Sec IV-B defines success as 'the user approaches, grasps, and lifts the object within 10 seconds.' During deployment the user drives the arm while HannesImitationPolicy generates wrist F/E, wrist P/S, and hand O/C actions. The reported 79.3% and 76% success rates therefore measure a joint human-policy system, not the policy alone. The user can compensate for poor or delayed wrist predictions by repositioning the whole arm, and the paper provides no ablation or control condition that removes or perturbs the policy's contribution. Offline action errors (Sec IV-A) and the qualitative traces in Figs. 4-6 show the policy produces sensible motions, but they do not establish that the high online success is attributable to those motions rather than to the user's arm-level compensation. Without such an attribution check, the central claim that a single diffusion policy can control the prosthesis across varied objects is underdetermined by the evaluation protocol.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents HannesImitation, an imitation-learning pipeline for the Hannes prosthetic hand. It introduces a dataset of 450 demonstrations (15 YCB objects, three scenarios: table grasp, shelf grasp, and human-to-prosthesis handover) collected with keyboard teleoperation, and adapts Diffusion Policy to predict wrist flexion/extension, wrist pronation/supination, and hand open/close from an eye-in-hand camera and joint encoders. The policy is deployed on the physical Hannes hand, and the authors report a 79.3% success rate over 450 trials on trained objects and 76% on five unseen YCB objects. They also compare with a visual-servo wrist controller [19] and claim an average improvement of 13.8%. The paper further states that the HannesImitationDataset is released to foster imitation-learning research for prosthetics.","tokens_in":12189,"tokens_out":4523,"duration_ms":45979,"significance":"The work has several genuine strengths: it involves physical deployment with a substantial number of real-robot trials (450 in-distribution plus 150 unseen-object trials), uses standard YCB objects across multiple scenarios, and appears to be the first application of a diffusion policy to a prosthetic hand with eye-in-hand vision. The claimed dataset release is a useful community contribution. If the results are robust, the paper is a meaningful step toward reducing cognitive load in prosthetic grasping. However, the central claim is currently underdetermined because the online success metric includes the human operator’s arm-level positioning, the statistical evidence lacks confidence intervals and significance tests, and the baseline comparison uses a different success criterion. These issues are substantive but addressable, and I believe they can be fixed with additional experiments and analysis.","major_comments":[{"comment":"The success metric is defined as 'the user approaches, grasps, and lifts the object within 10 seconds,' so the reported 79.3% and 76% rates measure a joint human-policy system rather than the policy alone. The user can compensate for poor or delayed wrist predictions by repositioning the whole arm, and the paper provides no control condition that removes or perturbs the policy’s contribution, nor any analysis linking offline action errors (Fig. 3) to online outcomes. I recommend adding a no-policy baseline (e.g., the user directly controls all three DoFs), an ablation with wrist or hand predictions disabled, or a per-trial analysis correlating predicted action quality with success; without such an attribution check, the central claim that 'a single diffusion policy' controls the prosthesis is not established by the deployment experiments.","section":"§IV-B, Table II"},{"comment":"All success rates are proportions from 10 trials per object with no confidence intervals, error bars, or significance tests. For example, in Table II the 044 flat screwdriver achieves 1/10 and the 065-g cups achieve 2/10 in the Shelf Grasp, yet the aggregate comparison treats 68% versus 80.6% as meaningful. Likewise, the headline claim that HannesImitation outperforms the visual servo by 13.8% in Table IV is based on averages with no statistical test, and per scenario the visual servo is better in two of the three tasks. Please report binomial confidence intervals (e.g., Clopper–Pearson) and appropriate tests (e.g., Fisher’s exact or a permutation test) for the key comparisons.","section":"§IV-A–IV-D, Tables II and III"},{"comment":"The comparison with the visual-servo baseline is not apples-to-apples. For the visual servo [19], success is defined as correct segmentation and tracking plus correct final wrist configuration, not as actually grasping and lifting the object, whereas HannesImitation success requires a physical grasp-and-lift. Additionally, in the baseline the user triggers the prediction and closes the fingers via EMG, while HannesImitation controls finger closure automatically, and the segmentation model in [19] was trained on the five test objects. These differing success criteria and unequal conditions make the statement 'our method outperforms [19] by 13.8%' unsupported as stated. I suggest re-evaluating [19] with the same grasp-and-lift criterion, or clearly labeling the comparison as a component-level measure rather than a task-level comparison.","section":"§IV-D, Table IV"},{"comment":"The deployment is performed by operators whose characteristics are not described, and the conclusion explicitly lists 'conducting user studies to assess real-world usability and impact' as future work. Since the human operator is part of the evaluated loop and the target population is trans-radial amputees, the current experiments cannot substantiate claims about practical prosthetic use or cognitive-load reduction. This is not necessarily a fatal flaw for a feasibility study, but the abstract and conclusion should scope the claims to the tested operator population, and this limitation should be stated in the main text rather than only in the future-work section.","section":"§IV-B and §V"}],"minor_comments":[{"comment":"Typo: 'obtaines' should be 'obtains' in the sentence reporting the overall success rate.","section":"§IV-B"},{"comment":"The paper states that the HannesImitationDataset is released, but no download link or availability statement appears in the manuscript; please add one.","section":"§III-B"},{"comment":"The box plots would be easier to interpret if the caption defined the box and whisker elements and stated the exact normalization range used for the absolute action error.","section":"Fig. 3"},{"comment":"The average success rates for both methods (76% and 62.6%) appear only in the text; adding them to the table would improve clarity.","section":"Table IV"},{"comment":"The handover experiments mention five subjects, but the manuscript gives no demographic or procedural details about these participants or about the operators who drove the prosthesis during data collection and testing.","section":"§III-B and §IV-B.3"}],"recommendation":"major_revision","confidential_remarks":"The paper comes from the same group as the visual-servo baseline [19], and the comparison is used as a central selling point; I would ask the editor to ensure that the revised version treats this comparison even-handedly. The ethical approval is stated, but the manuscript is vague about who operated the prosthesis and how many operators were involved; this should be clarified in revision. The novelty over [19] is real but incremental: the diffusion-policy module replaces the segmentation/visual-servoing module, while the human-in-the-loop evaluation makes the performance gain difficult to quantify. With the requested attribution checks and statistical reporting, the paper would be a solid contribution to the prosthetics and imitation-learning communities."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this is a careful, honest application of Diffusion Policy to prosthetic grasping, with a new dataset that will be useful to the field. The central feasibility claim—that a single learned policy can drive wrist and hand closure from an eye-in-hand camera across table, shelf, and handover scenarios—is supported by 450 real-robot trials. I'd send it to review.\n\nWhat's actually new: the HannesImitationDataset, the first IL dataset for prosthetic hands (450 demos, 15 YCB objects, three scenarios), and the demonstration that DP transfers to this low-cost, high-frequency setting. The 35 Hz inference on a laptop is a real practical point. The paper also does the right thing by testing on five unseen objects and showing comparable performance (76% vs 79.3%). That's a genuine generalization result.\n\nSoft spots, in order of importance. First, the success metric is a joint human-policy system. The user drives the arm; success means the user approaches, grasps, and lifts within 10 seconds. The policy controls wrist and fingers, but an able-bodied user can compensate for bad wrist predictions by moving the whole arm. Without a no-policy baseline or a perturbation test, the 79.3% doesn't cleanly attribute to the learned actions. The offline error plots are suggestive but not decisive. This is a real gap, not a fatal flaw.\n\nSecond, no statistics. Ten trials per object gives you a standard error of roughly 15 percentage points; the per-object numbers range from 1/10 to 10/10. A binomial confidence interval would show the aggregate 79.3% is somewhere in the 75-83% band, and the per-object differences are mostly noise. The paper should report intervals or at least raw counts (it does report counts, which is good, but no inference).\n\nThird, the visual servo comparison is not apples-to-apples. The baseline's success definition is whether segmentation and final wrist configuration are correct, not whether a grasp occurred. That's a different quantity. The paper acknowledges the baseline was trained on the test objects, but the metric mismatch undercuts the 'outperforms by 13.8%' claim. The handover result (86% vs 16%) is still striking, but I'd want the same metric on both.\n\nThe absent amputee testing is a limitation, but the paper says so clearly in future work. I don't hold that against it for a conference-rate result.\n\nBottom line: worth a serious referee. The dataset and the real-robot experiments are contributions; the attribution and statistics issues are fixable in revision.\n\nRecommendation: engage, and push for the no-policy baseline and CIs before publication.","headline":"Solid empirical IL-for-prosthetics paper with a useful dataset; the core feasibility claim holds, but the evaluation doesn't isolate the policy from the user and the baseline comparison is not apples-to-apples.","tokens_in":12687,"tokens_out":2357,"would_cite":false,"duration_ms":22118,"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":"Imitation learning drives a prosthetic hand to 79.3% grasp success across 450 trials.","keywords":["imitation learning","prosthetic hand","diffusion policy","grasping","human-to-prosthesis handover","eye-in-hand camera","behavior cloning","YCB objects"],"falsifier":"Run the same 10-trials-per-object protocol with trans-radial amputee users operating the Hannes hand through their customary socket and myoelectric interface; if their aggregate success falls to the level of a fixed-wrist baseline that simply closes the hand on approach, the claim that imitation learning enables practical prosthetic grasping would be refuted.","tokens_in":11823,"feed_emoji":"🦾","tokens_out":8327,"duration_ms":73679,"temperature":0.7,"pith_summary":"This paper claims that a prosthetic hand can learn to grasp everyday objects from demonstrations, rather than from hand-engineered vision rules. The authors train one diffusion-based visuomotor policy on 450 recorded grasps with the Hannes prosthetic hand, covering table-top grasps, shelf grasps, and human-to-prosthesis handovers. Deployed on the physical hand, the policy reads a palm-mounted camera and joint encoders, then commands wrist orientation and finger closure; it succeeds in 79.3% of 450 trials on the trained objects and 76% of 150 trials on five unseen objects. The point of the work is to show that shared-autonomy prosthetic grasping can be built from behavior cloning, potentially reducing the user's per-joint cognitive load and removing the need for labeled training data.","feed_headline":"Imitation learning drives a prosthetic hand to 79.3% grasp success","feed_subtitle":"One trained model controls wrist and fingers from a palm camera across table, shelf, and handover tasks","key_machinery":"The load-bearing object is the diffusion policy: a visuomotor policy that generates actions by iteratively denoising random noise conditioned on the current observation. Here it is trained with the standard denoising-diffusion objective, predicting the noise added to ground-truth action sequences, on a new dataset of 450 keyboard-driven demonstrations of the Hannes hand grasping 15 YCB objects in three scenarios. The observation encoder combines a ResNet-18 trained from scratch on palm-camera images with proprioceptive encoder readings, while the action decoder is a compact 1D temporal convolutional U-Net, and inference runs 10 denoising iterations per control step. This generative formulation is what lets one policy represent several valid approach and wrist behaviors for the same object while remaining stable to train, which is the mechanism that carries the generalization claim.","core_discovery":"On the paper's own terms, the discovery is that a single diffusion policy, adapted to the three actuated degrees of freedom of the Hannes hand, can act as the grasping controller across varied objects and environments. The policy takes the latest two palm-camera frames plus encoder readings for hand opening and wrist flexion, and outputs an eight-step action sequence covering hand open/close, wrist flexion/extension, and wrist pronation/supination; the hand executes four steps and then replans, running at about 35 Hz. In physical deployment it reaches 80.6% success for table grasps, 68% for shelf grasps, and 89.3% for human-to-Hannes handovers, a combined 79.3% over 450 trials, and 76% over 150 trials on objects never seen in training. On the handover scenario it clearly beats a segmentation-based visual servo controller, which confuses the object with the person's body; this is offered as evidence that demonstration learning generalizes better than manually labeled vision pipelines in unconstrained settings.","pith_inferences":["If the approach transfers to amputee users, the keyboard-driven demonstration pipeline could make data collection much cheaper than teleoperation, since demonstrations are recorded directly on the prosthesis without extra robot hardware.","The handover result hints that the same imitation-learning stack could be extended to socially contingent actions, such as waiting for the other person's grip to relax or releasing the object on cue, by adding demonstrations of those transitions, something the paper does not attempt.","A testable extension would be to perturb the object's pose or the user's approach velocity during deployment; the paper randomizes initial poses at data-collection time, so its robustness to larger perturbations remains unmeasured.","The wrist pronation/supination joint is velocity-controlled and lacks a position encoder, and the paper reports slightly higher action error for it; adding position feedback to that joint is a concrete candidate for closing the gap between shelf-grasp success (68%) and the other two scenarios."],"forward_implications":["A single trained policy can take over both wrist orientation and finger closure during a grasp, so the user no longer has to command each joint separately, which is the shared-autonomy benefit the paper aims at.","The small drop from 79.3% on trained objects to 76% on five unseen objects suggests the demonstration set and diffusion formulation transfer to novel objects without fine-tuning.","Grasping an object held by another person is learnable from demonstrations and is more robust than segmentation-based visual servoing in that scenario, because the policy does not need to separate the object from the person's body.","One policy, rather than per-object or per-scenario controllers, covers table, shelf, and handover contexts, and the released dataset allows other groups to reproduce or extend the result.","The roughly 35 Hz inference rate on a laptop GPU is compatible with closed-loop prosthetic use, making the approach feasible outside a dedicated robotics lab."],"supporting_citations":[{"why":"describes the Hannes prosthetic hand, the device the policy is trained on and deployed to","marker":"[4]"},{"why":"the segmentation-based visual servo wrist controller used as the comparison baseline in the unstructured-scenario experiments","marker":"[19]"},{"why":"introduces the Diffusion Policy method that the paper adapts for prosthetic grasping","marker":"[20]"},{"why":"defines the three-degree-of-freedom wrist and hand setup of the Hannes prosthesis used in this work","marker":"[41]"},{"why":"supplies the YCB object set from which the 15 training and 5 unseen objects are drawn","marker":"[42]"},{"why":"provides the ResNet-18 architecture used as the image feature extractor in the observation encoder","marker":"[43]"}],"fun_headline_variants":["Imitation learning teaches prosthetic hand to grasp diverse objects","One diffusion policy drives prosthetic grasps from palm camera","Prosthetic hand beats visual servo with imitation-learned policy","79% grasp success for imitation-driven prosthetic hand","Diffusion policy controls Hannes hand for real-world grasping"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that grasp performance measured with the research team operating the prosthesis transfers to actual amputee users, whose residual-limb control, arm kinematics, and cognitive load differ from the non-amputee, keyboard-demonstrated setup evaluated here.","fun_headline_variants_meta":{"raw":{"variants":["Imitation learning teaches prosthetic hand to grasp diverse objects","One diffusion policy drives prosthetic grasps from palm camera","Prosthetic hand beats visual servo with imitation-learned policy","79% grasp success for imitation-driven prosthetic hand","Diffusion policy controls Hannes hand for real-world grasping"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000662,"raw_usage":{"total_tokens":3048,"prompt_tokens":993,"completion_tokens":2055,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":609,"completion_tokens_details":{"reasoning_tokens":1977}},"tokens_in":609,"tokens_out":2055,"duration_ms":14519,"temperature":1.0,"reasoning_tokens":1977,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T10:06:28.708479+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same 10-trials-per-object protocol with trans-radial amputee users operating the Hannes hand through their customary socket and myoelectric interface; if their aggregate success falls to the level of a fixed-wrist baseline that simply closes the hand on approach, the claim that imitation learning enables practical prosthetic grasping would be refuted.","supporting_citations":[{"cited_title":"The hannes hand prosthesis replicates the key biological properties of the human hand,","cited_arxiv_id":null,"evidence_quote":"describes the Hannes prosthetic hand, the device the policy is trained on and deployed to"},{"cited_title":"Continuous wrist control on the hannes prosthesis: a vision-based shared autonomy framework,","cited_arxiv_id":null,"evidence_quote":"the segmentation-based visual servo wrist controller used as the comparison baseline in the unstructured-scenario experiments"},{"cited_title":"Diffusion policy: Visuomotor policy learning via action diffusion,","cited_arxiv_id":null,"evidence_quote":"introduces the Diffusion Policy method that the paper adapts for prosthetic grasping"},{"cited_title":"Development of a 2-dofs actuated wrist for enhancing the dexterity of myoelectric hands,","cited_arxiv_id":null,"evidence_quote":"defines the three-degree-of-freedom wrist and hand setup of the Hannes prosthesis used in this work"},{"cited_title":"Benchmarking in manipulation research: Using the yale- cmu-berkeley object and model set,","cited_arxiv_id":null,"evidence_quote":"supplies the YCB object set from which the 15 training and 5 unseen objects are drawn"}],"review_version":1}