{"id":"9ca2d39c-c263-4290-bec3-eb3c8806a82e","arxiv_id":"2501.11742","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Adding force measurements to an imitation-learning policy improves autonomous tissue retraction success on a dVRK robot, though the reported gentleness advantage is much weaker on successful trials alone.","lead":"This paper compares two imitation-learning policies for autonomous tissue retraction on a surgical robot, one with tool-tissue force measurements and one without. The force-aware policy completed the task far more often, but its reported gentleness advantage mostly disappears when only successful trials are counted.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The stiffness-generalization claim is not identifiable: the two silicone samples differ in material composition, not only Young's modulus, so the unseen-sample result may reflect friction or surface differences rather than stiffness.","rationale":"The reader's weakest_assumption identifies the same confound, so I agree. This is the most load-bearing concern because the paper's motivating hypothesis is specifically about manipulating tissues of different stiffness levels; if the two training/test samples are not stiffness-matched isolates, the key generalization result is uninterpretable. The reader's secondary point about gentleness is also valid: comparing pooled all-rollout means mixes success and failure, and successful-rollout means are close (0.26 vs 0.28 N seen; 0.39 vs 0.42 N unseen), so the 'gentler' claim is overstated. But the stiffness confound is more fundamental to the paper's stated contribution. Both issues are already captured by a conditional verdict, so I leave the reader's verdict unchanged.","tokens_in":12386,"tokens_out":7709,"duration_ms":86905,"concrete_test":"Fabricate at least four tissue samples spanning the same 55-151 kPa modulus range while keeping surface chemistry and friction roughly constant, e.g., by varying the mixing ratio of one Ecoflex or Dragon Skin base or by using a single base with different amounts of silicone thinner. Characterize each sample's 100% modulus, coefficient of friction, and surface tack. Rerun the 50-rollout protocol for both the force and no-force policies on each unseen sample. If the force policy's success and force advantage tracks Young's modulus while friction and tack are held constant, the stiffness-generalization claim is supported; if the advantage appears only when friction or tack change, the claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that force observations enable a policy to generalize to tissues of different stiffness (hypothesis in Section III.E, conclusion in Section VI) rests on a single two-sample comparison. Section III.C states that the less stiff and more stiff samples are made from two different Smooth-On materials, Dragon Skin and Ecoflex, with 100% modulus 55 kPa and 151 kPa. These materials differ in many properties besides Young's modulus, including coefficient of friction, surface tack, tear strength, and density. Because there is only one sample per material, the force policy's 70% vs 20% success on the unseen sample cannot be attributed to stiffness adaptation; it could equally be adaptation to a different surface-friction or tear regime. The paper honestly discloses the materials, but the experimental design does not isolate the variable named in the hypothesis. If this confound is real, the 'generalize to an unseen tissue sample of different stiffness level' contribution is not established, although the seen-tissue success benefit and the raw transfer to a different silicone remain.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper trains two Action-Chunking Transformer (ACT) imitation-learning policies for autonomous tissue retraction on the da Vinci Research Kit: a 'force policy' that observes tool-tissue force/torque in addition to stereo images and joint positions, and a 'no force policy' that observes only images and joint positions. Both policies are trained on 60 demonstrations from the same silicone tissue sample and evaluated with 50 rollouts each on the training (seen) sample and on a second, stiffer silicone (unseen) sample. The authors report that the force policy achieves higher success (76% vs 26% on seen tissue, 70% vs 20% on unseen tissue) and lower mean applied force over all rollouts (0.29 vs 0.47 N and 0.40 vs 0.84 N, respectively). They interpret these results as evidence that force-aware autonomous systems are more successful, gentler, and better able to generalize across tissue stiffness levels.","tokens_in":12578,"tokens_out":6280,"duration_ms":60242,"significance":"If the results hold, the paper provides a clear and practically useful demonstration that force/torque observations can improve imitation-learning policies for a contact-rich surgical subtask. The experimental setup uses a real surgical robot, physical tissue phantoms, and a systematic force/no-force ablation, and the success-rate differences are large and consistent across seen and unseen samples. However, the generalization-to-stiffness claim is confounded by the use of two different silicone materials, and the gentleness claim is substantially weaker when only successful rollouts are compared. These issues limit the strength of the stated conclusions but do not undermine the core finding that force input improves task success in this setup.","major_comments":[{"comment":"The stiffness-generalization claim is not identifiable from the current experimental design. Section III.C states that the two tissue samples were made from two different silicone materials (Dragon Skin and Ecoflex) with 100% moduli of 55 kPa and 151 kPa. These materials differ in properties beyond Young's modulus, including surface friction, tack, tear strength, and density, and there is only one sample per material. The higher success of the force policy on the unseen sample (70% vs 20% in Table III) could therefore reflect adaptation to a different surface-friction or tear regime rather than to tissue stiffness, which is the variable named in the hypothesis (Section III.E) and the conclusion (Section VI). To support the stiffness claim, the authors should use samples of the same material with controlled stiffness differences, or measure and rule out changes in other mechanical properties, and ideally include multiple samples per condition.","section":"Section III.C and Section VI"},{"comment":"The gentleness claim relies on pooled force means over all rollouts, but the comparison on successful rollouts only shows small differences (0.26 vs 0.28 N on seen tissue, 0.39 vs 0.42 N on unseen tissue), whereas the all-rollout means differ much more (0.29 vs 0.47 N and 0.40 vs 0.84 N). Because the no-force policy fails more often, the lower all-rollout mean force may be driven by failure episodes rather than by gentler interaction during successful task execution. The paper should report the successful-rollout comparison as the primary evidence for gentleness, or provide an analysis that separates the effect of success from the effect of force regulation during execution.","section":"Section IV, Tables II and III"},{"comment":"The abstract claims that on the unseen tissue sample the force policy exerts 'an order of magnitude less force' than the no-force policy, but Table III reports mean forces of 0.40 N vs 0.84 N, which is a factor of about 2.1, not 10. The body text correctly states that the no-force policy applies 110% more force on average. The abstract should be corrected to be consistent with the data.","section":"Abstract and Section IV.B"},{"comment":"The success rates (76% vs 26% and 70% vs 20%) are reported without confidence intervals or a significance test, and the t-tests for force means are not described in sufficient detail. If the t-tests were performed on pooled time-series samples, the independence assumption is violated and the effective sample size is inflated, making the reported p-values (p < 0.01) uninformative. The authors should either provide per-rollout summaries (e.g., mean force per rollout) with appropriate tests, or account for temporal correlation, and include confidence intervals for the success-rate differences.","section":"Section IV.A and IV.B"}],"minor_comments":[{"comment":"The averaging window size for force smoothing is not specified; please report it.","section":"Section III.D"},{"comment":"The text does not state which silicone material corresponds to which modulus (55 kPa vs 151 kPa); specifying this would help readers assess the material confound.","section":"Section III.C"},{"comment":"The y-axis label 'Normalized Time [s]' is unclear; normalized duration should be dimensionless or the normalization should be described in the caption.","section":"Figures 4 and 5"},{"comment":"'H-params' should be written as 'Hyper-parameters'.","section":"Table I"},{"comment":"The discussion attributes contact detection to force data, but the force/torque sensor is mounted beneath the tissue, not at the tool-tissue interface; the text should clarify this distinction.","section":"Section V"}],"recommendation":"major_revision","confidential_remarks":"The central ablation is a reasonable empirical study, but the stiffness-generalization claim needs additional experiments. The shared authorship with reference [43] is not, in my view, a circularity problem, since the comparison does not depend on that work's predictions; however, the paper may need an explicit disclosure if the journal requires it. The paper fits the scope of the journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: the headline result — force-conditioned ACT beats force-agnostic ACT on tissue retraction — is supported by the data as a same-task ablation. Fifty rollouts per policy give 76% vs 26% success on seen tissue and 70% vs 20% on the unseen sample. That is a large, consistent gap. The paper is the first to run this particular force/no-force comparison for learned autonomous retraction on the dVRK, and it is a useful baseline for anyone building surgical imitation-learning systems.\n\nThe main soft spot is the unseen-tissue comparison. The less stiff and more stiff samples are made from Dragon Skin and Ecoflex, two different Smooth-On materials. With one sample per material, you cannot attribute the force policy's 70% success to stiffness adaptation; it could be friction, surface tack, or tear behavior. The paper honestly names the materials and the hypothesis in Section III.E, but the conclusion in Section VI overreaches the design. The seen-tissue success benefit stands; the stiffness-generalization claim needs multiple samples per stiffness level or material characterization.\n\nThe gentleness claim is also overstated. The all-rollout pooled force means differ (0.29 vs 0.47 N seen; 0.40 vs 0.84 N unseen), but pooling over rollouts includes failures, where the no-force policy may press into the tissue while failing. On successful rollouts only, the means are nearly identical (0.26 vs 0.28 N seen; 0.39 vs 0.42 N unseen). The t-test on force time series is not meaningful because those samples are not independent. The honest claim is \"more successful, and lower average force over all attempts,\" not \"gentler.\" Max forces are similar between policies.\n\nMinor issues: no confidence intervals or significance test for the success proportions, and no code or data released. Neither is fatal; the success gap is large and the setup is described well enough to reproduce. The shared authorship with the force-estimation paper [43] is fine because the central ablation does not depend on that work.\n\nThe paper deserves a serious referee. I would send it out, but ask for revision: reframe or properly test the stiffness claim, report successful-rollout force statistics, and ideally release the data. It is an honest, useful empirical contribution, just less strong than its own conclusions claim.","headline":"Genuinely useful force/no-force ablation for learned surgical retraction, but the stiffness-generalization claim overreaches a two-material, one-sample comparison.","tokens_in":13098,"tokens_out":2573,"would_cite":true,"duration_ms":28145,"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":"Feeding force measurements into an imitation-learning policy tripled autonomous tissue-retraction success and made the robot gentler.","keywords":["autonomous robotic surgery","imitation learning","action-chunking transformer","force-aware policy","tissue retraction","tool-tissue interaction forces","generalization"],"falsifier":"Make a family of tissue samples from a single silicone base with stiffness varied continuously by mixing ratio and identical surface finish, then roll both policies out across that range; if the force policy's advantage does not track stiffness, the claim that force input enables stiffness generalization is not supported.","tokens_in":12222,"feed_emoji":"🩺","tokens_out":7040,"duration_ms":68905,"temperature":0.7,"pith_summary":"Autonomous surgical robots have to touch tissue at different stiffnesses, and this paper asks whether giving the robot force feedback lets it do that well. The authors compare two imitation-learning policies for a tissue-retraction task on a surgical robot, both built from the same 60 human demonstrations: one receives six-axis tool-tissue force/torque data along with stereo video and joint angles, the other receives only video and joint angles. On the training tissue, the force-aware policy succeeded 76% of the time versus 26% for the force-agnostic one; on a stiffer, unseen tissue, it succeeded 70% versus 20%. The force-aware policy also applied less force, with the no-force policy averaging 62% more force on seen tissue and 110% more on unseen tissue. The intended significance is that force sensing could let autonomous systems meet surgical guidelines that call for tissue-appropriate handling.","feed_headline":"Adding force sensing triples autonomous surgical success rates","feed_subtitle":"A force-aware imitation-learning policy beat a vision-only one on success and applied force, on seen and unseen tissue.","key_machinery":"The load-bearing component is the modified action-chunking transformer, a conditional variational autoencoder whose decoder acts as the policy and predicts the next 100 joint actions. The modification adds six numbers—the forces and torques at the tool-tissue interface—to the observation the transformer sees, together with left/right camera images and the robot's joint angles. Those six numbers carry the argument because they give the policy a direct contact cue: they signal when the grasper has actually touched the tissue, which the paper argues is hard to infer from vision alone when the tissue is stiff and deforms little. The force input also encodes how hard the demonstrator pulled, giving the policy a target force profile to imitate.","core_discovery":"The central claim is that tool-tissue interaction force is a decisive input modality for imitation-learned surgical manipulation, not a luxury. A modified action-chunking transformer—an imitation-learning architecture that predicts a fixed-length sequence of joint actions—accepts a six-dimensional force/torque vector alongside stereo images and joint positions. Trained on the same demonstrations, this force-aware policy becomes roughly three times more successful at retracting a tissue flap and applies markedly lower forces than the identical architecture without force input. The advantage persists on a previously unseen stiffer tissue sample, which the paper takes as evidence that force input supports generalization across tissue stiffness. The paper also reports the force-aware policy's force profile is smoother, with most applied forces below 0.5 N on the unseen sample.","pith_inferences":["The paper's pooled force numbers mix successful and failed rollouts, so part of the force reduction may reflect fewer failed grasps rather than gentler execution; isolating successful rollouts would separate 'less fumbling' from 'more delicate touch.'","The same contact-cue mechanism should transfer to other contact-rich subtasks like suturing or dissection, but the paper does not test that.","A natural next experiment is to replace the physical force sensor with a vision-based force estimator; if the benefit survives estimated forces, the approach becomes deployable on the large installed base of surgical robots without force sensing."],"forward_implications":["If the claim is right, any surgical robot that can sense tool-tissue forces should include those forces in imitation-learning policies; in this study the same architecture became roughly three times more successful and gentler with force input than without.","The force policy's performance on a stiffer unseen sample (70% versus 20%) suggests force-aware policies can transfer to tissue stiffnesses not present in the demonstration data without retraining.","The force gap—no-force policy averaging 62% more force on seen tissue and 110% more on unseen tissue—indicates that force-aware execution is closer to the gentle handling surgical guidelines recommend.","The approach is portable: it works with a physical force sensor, and the paper argues it can be paired with vision-based force estimation on robots that lack force sensing."],"supporting_citations":[{"why":"Supplies the action-chunking transformer imitation-learning architecture that this paper modifies with a force/torque input.","marker":"[39]"},{"why":"Shows the same ACT architecture can imitate surgical subtasks on the da Vinci robot, the baseline method this paper builds on.","marker":"[38]"},{"why":"Defines the da Vinci Research Kit hardware used for teleoperated demonstrations and autonomous rollouts.","marker":"[50]"},{"why":"Defines tissue retraction and why it matters as a surgical subtask, the task this paper automates.","marker":"[8]"},{"why":"States surgical guidance that forces should match tissue characteristics, motivating the force-aware design.","marker":"[11]"},{"why":"Provides a vision-and-robot-state force estimation method the paper suggests as a substitute for physical force sensing on clinical robots.","marker":"[43]"}],"fun_headline_variants":["Force sensing triples autonomous surgical success","Force-aware robot surgery: 3x success, 62% less force","Adding force input triples surgical robot success","Force feedback boosts robotic surgery success by 3x","Force-aware imitation learning triples surgery success"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The generalization result assumes the two silicone tissue samples differ only in stiffness, but they are made from different silicone formulations, so friction, surface texture, or tear behavior could also explain why the force policy did better on the unseen sample.","fun_headline_variants_meta":{"raw":{"variants":["Force sensing triples autonomous surgical success","Force-aware robot surgery: 3x success, 62% less force","Adding force input triples surgical robot success","Force feedback boosts robotic surgery success by 3x","Force-aware imitation learning triples surgery success"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00035,"raw_usage":{"total_tokens":1922,"prompt_tokens":965,"completion_tokens":957,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":581,"completion_tokens_details":{"reasoning_tokens":883}},"tokens_in":581,"tokens_out":957,"duration_ms":8918,"temperature":1.0,"reasoning_tokens":883,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T17:54:46.039922+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Make a family of tissue samples from a single silicone base with stiffness varied continuously by mixing ratio and identical surface finish, then roll both policies out across that range; if the force policy's advantage does not track stiffness, the claim that force input enables stiffness generalization is not supported.","supporting_citations":[{"cited_title":"Learning fine-grained bimanual manipulation with low-cost hardware,","cited_arxiv_id":null,"evidence_quote":"Supplies the action-chunking transformer imitation-learning architecture that this paper modifies with a force/torque input."},{"cited_title":"Surgical robot transformer (srt): Imitation learning for surgical subtasks,","cited_arxiv_id":null,"evidence_quote":"Shows the same ACT architecture can imitate surgical subtasks on the da Vinci robot, the baseline method this paper builds on."},{"cited_title":"An open-source research kit for the da Vinci Surgical System,","cited_arxiv_id":null,"evidence_quote":"Defines the da Vinci Research Kit hardware used for teleoperated demonstrations and autonomous rollouts."},{"cited_title":"Current and future practices in surgical retraction,","cited_arxiv_id":null,"evidence_quote":"Defines tissue retraction and why it matters as a surgical subtask, the task this paper automates."},{"cited_title":"Human vs robotic organ retraction during laparoscopic nissen fundoplication,","cited_arxiv_id":null,"evidence_quote":"States surgical guidance that forces should match tissue characteristics, motivating the force-aware design."},{"cited_title":"Toward force estimation in robot-assisted surgery using deep learning with vision and robot state,","cited_arxiv_id":null,"evidence_quote":"Provides a vision-and-robot-state force estimation method the paper suggests as a substitute for physical force sensing on clinical robots."}],"review_version":1}