{"id":"b19369cd-5a7f-4d54-95f7-12b5379fba06","arxiv_id":"2501.10356","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"DexForce extracts force-informed position targets from kinesthetic demonstrations, enabling imitation-learned policies to succeed on contact-rich dexterous tasks where position-only actions fail.","lead":"This paper introduces DexForce, a method that uses fingertip force sensors to turn human-guided robot demonstrations into actions that include the needed contact forces. Policies trained on these force-aware actions succeeded on 76% of trials across six contact-rich tasks, while policies using only finger positions nearly always failed.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central comparison lacks error bars and seed variance; 76% vs. near-zero may not be statistically robust.","rationale":"The reader's weakest_assumption focuses on the spring model in Eq. 3, but the sensitivity analysis (Fig. 9) already addresses this partially: it sweeps over kf for five demonstrations per task and shows a permissive range of values that all lead to successful replays, with the chosen kf comfortably inside the overlap. This empirical support weakens the spring-model concern for the claimed tasks. The statistical reliability of the central comparison, however, is not addressed anywhere in the paper. The main result is a single point estimate with no uncertainty quantification; this is particularly problematic because the paper demonstrates that the authors know how to perform significance testing (Fisher's exact test in Question 2) yet omits it for the primary claim. A single-seed Diffusion Policy run can be noisy, and the near-zero baseline could be a few lucky or unlucky trials away from a qualitatively different number. The 30-trial evaluation is also moderate in size, so paired tests are essential to account for shared initial configurations. Because the claim is specifically that force-informed actions yield 76% versus near-zero, the absence of error bars directly threatens the credibility of that number. I do not think this warrants rejection; the mechanism is plausible and the difference is large, so the appropriate outcome is still conditional acceptance pending these additional statistical details. The reader's verdict of CONDITIONAL is therefore unchanged, but the specific condition should be re-focused on reporting seed variance and significance tests for Question 1 rather than on the spring-model fidelity.","tokens_in":11311,"tokens_out":14217,"duration_ms":149136,"concrete_test":"Retrain each of the 12 policy conditions (six tasks times two action representations) with five random seeds, evaluate each seed on the same 30 initial configurations, and report mean success rate, standard error, and the observed distribution across seeds. For each task, apply McNemar's test to the paired 30-configuration outcomes (force-informed vs. position-only) to determine whether the improvement is statistically significant; also report the distribution of the average success rate across tasks and seeds.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline claim (Section IV, Question 1) rests on point estimates: force-informed-action policies average 76% success across six tasks, while position-only-action policies have 'near-zero' success. The paper reports no error bars, no number of training seeds, and no significance test for this primary comparison, yet it uses Fisher's exact test for the observation ablation in Question 2. Diffusion Policies are stochastic: a single seed per condition can yield success rates far from the mean, especially for a near-zero baseline where a few accidental successes change the rate qualitatively. Without seed variance or a paired statistical test that accounts for the shared 30 initial configurations, the paper cannot distinguish a systematic advantage of force-informed actions from training noise. This is the most load-bearing issue because the central claim is inherently quantitative. The spring-model assumption (Eq. 3) is partially mitigated by the sensitivity analysis (Fig. 9), which shows a broad permissive range of kf values and a minimum kf greater than zero, so the modeling premise is less fragile than the statistical basis of the headline comparison.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents DexForce, a demonstration-collection method for dexterous manipulation that converts kinesthetic demonstrations into force-informed actions. The action at each timestep is a fingertip position target x_f = x_o + k_f f, where x_o is the observed fingertip position and f is the measured contact force; tracking this target with a Cartesian impedance controller is intended to reproduce the demonstrated forces. Demonstrations are collected in two stages: first, force-informed targets are extracted from human kinesthetic manipulation; second, these targets are replayed by the robot to obtain robot-only image observations. Diffusion Policies are then trained on these demonstrations. Across six contact-rich tasks, policies trained on force-informed targets achieve an average 76% success rate, while a baseline trained on raw observed fingertip positions has near-zero success. Additional experiments ablate force observations and analyze sensitivity to the stiffness parameter k_f.","tokens_in":11495,"tokens_out":5196,"duration_ms":51090,"significance":"If the empirical claims hold, the paper makes a useful contribution: it demonstrates that the action representation itself, not just the observation, must encode contact forces for contact-rich dexterous manipulation, and it provides a simple, hardware-feasible way to compute such actions from kinesthetic demonstrations. The main experiment is internally consistent, and the ablation and sensitivity analysis (Figs. 8 and 9) are appropriate and informative. The k_f sensitivity study is a particular strength: it shows a permissive range of stiffness values across tasks, which supports the robustness of the spring-model approximation. The paper does not provide code or machine-checked proofs, but the method description is sufficiently clear for replication. The principal weakness is the lack of statistical reporting for the headline comparison, which currently makes the central quantitative claim under-supported.","major_comments":[{"comment":"The headline claim that force-informed actions achieve 76% average success while position-only actions achieve near-zero success is based on point estimates with no reported variance across training seeds and no significance test. Because Diffusion Policies are stochastic, a single seed per condition can yield success rates far from the mean, especially for a near-zero baseline where a few accidental successes qualitatively change the rate. Since all policies are evaluated on the same 30 initial configurations, a paired statistical test (e.g., McNemar's test) or confidence intervals over multiple training seeds for each task is needed to support the central quantitative claim. Please report the number of seeds, mean and standard deviation (or per-seed results), and a suitable test for the comparison in Fig. 7.","section":"Section IV, Question 1; Fig. 7"},{"comment":"The stiffness k_f was tuned (k_f = 0.0045) so that Stage-2 replays succeed, whereas the observed-position baseline in Question 1 has no comparable free parameter. This asymmetry makes the comparison in Fig. 7 less direct than it appears: the force-informed condition benefits from task-success-based parameter selection while the baseline does not. The sensitivity analysis in Fig. 9 shows a broad range of acceptable k_f values, which mitigates concerns about fragility, but the paper should explicitly discuss whether the position-only baseline is disadvantaged by lacking an equivalent tunable scale or offset, and ideally evaluate a position-only action with a simple tuned parameter (e.g., a constant offset or gain) to confirm that the advantage is due to force information rather than to the tuning procedure.","section":"Section III, Eq. (3); Section IV, Question 1"}],"minor_comments":[{"comment":"In the third observation-ablation bullet, 'RGB only: RBG image features only' contains a typo: 'RBG' should be 'RGB'.","section":"Section IV, Question 2"},{"comment":"Reference [12] is incomplete: 'arXiv preprint arXiv:240' is cut off and should include the full identifier.","section":"References"},{"comment":"For unscrew nut, the observed-position policy achieves 23/30 partial successes; please include partial-success rates in Fig. 7 or the text, because 'near-zero success' refers only to full successes and the partial successes provide important context for interpreting the baseline's behavior.","section":"Section IV, Question 1"},{"comment":"The scalar k_f is named similarly to the force vector f; consider using a different symbol (e.g., α) or explicitly stating 'scalar stiffness' at every use to avoid confusion.","section":"Section III, Eq. (3)"},{"comment":"The controller gains k_p and k_v are given as numbers without units; please specify their units or state that they are dimensionless gains for the Allegro hand's impedance control.","section":"Section III, Eq. (1)"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's scope and the central idea is worth publishing if the statistical concerns are addressed. The main risk is overclaiming from single-seed evaluations; please require the authors to either provide seed variance or temper the headline claim. The k_f tuning asymmetry with the baseline is a secondary concern that should be discussed explicitly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is new and the experiments mostly support it; the main gap is that the headline comparison (76% vs. near-zero) has no seed variance or error bars. Fix that and this is a solid paper.\n\nWhat's actually new: using measured contact forces to compute position targets for policies has been around for robot arms and for fixed-grasp hand manipulation, but DexForce is the first I know to apply it to multi-finger dexterous tasks that require contact switching, without per-task heuristics. The two-stage procedure (kinesthetic demonstration, then replay of force-informed targets to capture clean robot-only images) is a practical answer to the operator-in-frame problem. That alone makes the paper worth reading.\n\nIt does several things well. The observation ablation uses Fisher's exact test, which is more careful than most robot-learning papers. The sensitivity analysis of kf is honest and shows a broad operating range, so the spring model is not fragile in practice. The tasks are nontrivial and the failure analysis (e.g., partial successes on unscrew nut) shows real engagement with behavior. The related-work section covers arm-based and fixed-grasp hand precedents fairly; I didn't spot a missing obvious reference.\n\nThe soft spots are real but manageable. Question 1's comparison is point estimates only. Diffusion Policy is stochastic, and a single seed per condition makes 'near-zero' and '76%' hard to interpret. The paper should report seeds, means, and a paired test over the 30 shared initial configurations. No code or dataset release also limits reproducibility. The spring model ignores friction and soft contact, but the authors acknowledge it and the sensitivity analysis shows it is not the bottleneck.\n\nThis paper is for anyone collecting demonstration data for dexterous manipulation or studying the role of force information in policies. It deserves serious peer review; the central claim is plausible and the method is useful. The referee should demand statistical details rather than reject.","headline":"Useful new method for collecting contact-rich dexterous demonstrations; the headline comparison needs seed variance, though the effect likely holds.","tokens_in":12024,"tokens_out":2797,"would_cite":true,"duration_ms":29324,"reading_group":"yes","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 for contact-rich dexterous manipulation succeeds when actions are computed from measured contact forces, and the force-informed target $x_f = x_o + k_f f$ is the action representation that makes it work.","keywords":["force-informed actions","kinesthetic demonstration","dexterous manipulation","imitation learning","contact-rich manipulation","impedance control","force-torque sensing","diffusion policy"],"falsifier":"Take a kinesthetic demonstration of a task with significant tangential friction, such as sliding a cube while pressing down, and replay the force-informed target trajectory with $k_f = 0.0045$ while measuring the actual 6-axis force at the fingertip. If the reconstructed contact force directionally deviates from the demonstration's measured force and the deviation grows with the friction component, the single-scalar stiffness assumption fails exactly on tasks that need coordinated tangential forces.","tokens_in":11119,"feed_emoji":"🖐️","tokens_out":5885,"duration_ms":53565,"temperature":0.7,"pith_summary":"This paper argues that for contact-rich dexterous manipulation, the right way to define actions in imitation learning is to compute them from measured contact forces rather than from fingertip positions alone. It introduces DexForce, which converts kinesthetic demonstrations on an instrumented robot hand into force-informed position targets $x_f = x_o + k_f f$ by augmenting observed fingertip positions with a term proportional to the measured contact force. When these targets are tracked by an impedance controller and used to train diffusion policies, the resulting policies succeed on average 76% of the time across six tasks, while policies trained on raw fingertip positions have near-zero success. The paper also shows that adding 6-axis force data to policy observations never hurts and helps most on tasks that demand precise, coordinated contact, such as opening an AirPods case and unscrewing a nut.","feed_headline":"Force-aware actions lift dexterous policy success to 76%","feed_subtitle":"Policies trained on actions computed from fingertip contact forces beat position-only actions on six hand tasks.","key_machinery":"The load-bearing object is the force-informed target $x_f = x_o + k_f f$, a per-finger desired position computed by shifting the observed fingertip position $x_o$ along the measured contact force $f$ scaled by a scalar stiffness $k_f$. It converts a kinesthetic demonstration, which contains only states (positions and forces), into actions: when the Cartesian impedance controller tracks $x_f$, the first term of its control law, $F = k_p(x_d - x_c) - k_v \\dot{x}_c$, turns the displacement $x_f - x_c$ into an applied force. The identity that makes this work is the quasi-static spring relation between commanded position error and contact force, so the same $k_f$ is used in both demonstration replay (Stage 2) and policy runtime. The paper's sensitivity analysis shows a whole interval of $k_f$ values succeeds for each task and these intervals overlap, so the parameter does not need task-specific tuning.","core_discovery":"Measured contact forces from kinesthetic demonstrations contain information that observed positions do not: the same fingertip pose can coincide with very different applied forces. DexForce extracts this information into the action space by computing a force-informed target for each finger, $x_f = x_o + k_f f$, where $x_o$ is the recorded fingertip position, $f$ the measured contact force exerted by the finger on the object, and $k_f$ a single hand-tuned stiffness. The claim is that commanding the impedance controller to track $x_f$ reproduces both the motion and the force profile of the demonstration, and that policies supervising on these targets learn contact-rich skills that policies supervising on observed positions cannot. The evidence is a six-task evaluation on an Allegro hand with fingertip force-torque sensors: force-informed action policies average 76% success, position-only policies average near zero, and 6-axis force observations give the largest gains on the most force-sensitive tasks, with binary contact signals performing no better than no force signal at all.","pith_inferences":["If the 76% versus near-zero gap is robust, the dominant failure mode in teleoperation-based data collection for dexterous hands may be the action space itself rather than state coverage; adding force-shifted targets to existing retargeting pipelines could inherit this benefit at scale.","The single-scalar spring model suggests a testable generalization: replacing $k_f$ with a per-contact compliance that depends on normal direction or contact area should improve force reconstruction on soft or frictional contacts, and a principled way to select $k_f$ from measured kinematics would remove the main tuning knob.","The correlation the paper observes between task force-sensitivity and the benefit of force observations could be turned into a predictor: tasks with small admissible $k_f$ ranges are the ones where force observations matter most, so a short replay sweep could estimate the expected observation gain for a new task.","DexForce's two-stage procedure still requires a human operator; a semi-automated variant that replaces Stage 1 kinesthetic examples with scripted or RL-generated force profiles, keeping only the replay stage, would test whether the force-informed action representation alone, rather than the human demonstration, is what drives success."],"forward_implications":["Force-informed action extraction is what makes imitation learning succeed on contact-rich dexterous tasks; position-only actions collapse to near-zero success, so the action representation, not just the observation, determines learnability.","Including 6-axis force and moment data in policy observations is never harmful and yields large gains on tasks requiring precision and coordination; binary contact signals do not capture what is needed.","Because the admissible $k_f$ intervals overlap across all six tasks, a single stiffness parameter suffices for the whole suite, and DexForce does not require per-task tuning.","Replaying force-informed targets preserves enough of the demonstration's force profile to generate successful robot-only demonstrations, making the two-stage collection procedure with wrist-camera images feasible.","DexForce extends to tasks with contact switching and no fixed grasp, which earlier kinesthetic-teaching force methods could not handle without task-specific heuristics."],"supporting_citations":[{"why":"Supplies the diffusion policy architecture that consumes the DexForce actions as supervision targets.","marker":"[3]"},{"why":"Defines the Cartesian impedance control law whose position-error term converts $x_f$ into applied force.","marker":"[40]"},{"why":"Provides the coin-sized 6-axis force-torque fingertip sensors used to record contact forces during demonstrations.","marker":"[43]"},{"why":"Shows that binary contact signals can suffice for some dexterous tasks, forming the baseline the observation ablation must outperform.","marker":"[45]"},{"why":"Earlier kinesthetic-teaching force methods that assume a fixed non-slipping grasp, which DexForce generalizes beyond.","marker":"[26, 27]"},{"why":"Demonstrate force-informed imitation learning on robot arms, the principle DexForce transfers to dexterous hands.","marker":"[32, 33]"}],"fun_headline_variants":["Force-aware actions from demos lift success to 76%","Position-only actions near zero: force is key","DexForce: contact forces make dexterous policies learn","Forces during demos unlock contact-rich manipulation skills","Why force matters: 76% vs near zero on hand tasks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire pipeline rests on the quasi-static spring model $x_f = x_o + k_f f$: contact forces are assumed proportional to commanded position error through one scalar stiffness, ignoring soft contact and friction; if real contacts do not follow this relation, the force-informed targets will not reproduce the demonstration's forces.","fun_headline_variants_meta":{"raw":{"variants":["Force-aware actions from demos lift success to 76%","Position-only actions near zero: force is key","DexForce: contact forces make dexterous policies learn","Forces during demos unlock contact-rich manipulation skills","Why force matters: 76% vs near zero on hand tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000692,"raw_usage":{"total_tokens":3134,"prompt_tokens":949,"completion_tokens":2185,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":2103}},"tokens_in":565,"tokens_out":2185,"duration_ms":16595,"temperature":1.0,"reasoning_tokens":2103,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T19:09:20.020292+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a kinesthetic demonstration of a task with significant tangential friction, such as sliding a cube while pressing down, and replay the force-informed target trajectory with $k_f = 0.0045$ while measuring the actual 6-axis force at the fingertip. If the reconstructed contact force directionally deviates from the demonstration's measured force and the deviation grows with the friction component, the single-scalar stiffness assumption fails exactly on tasks that need coordinated tangential forces.","supporting_citations":[{"cited_title":"Impedance control: An approach to manipulation,","cited_arxiv_id":null,"evidence_quote":"Defines the Cartesian impedance control law whose position-error term converts $x_f$ into applied force."},{"cited_title":"Coinft: A coin-sized, capacitive 6-axis force torque sensor for robotic applications,","cited_arxiv_id":null,"evidence_quote":"Provides the coin-sized 6-axis force-torque fingertip sensors used to record contact forces during demonstrations."},{"cited_title":"Rotating without seeing: Towards in-hand dexterity through touch,","cited_arxiv_id":null,"evidence_quote":"Shows that binary contact signals can suffice for some dexterous tasks, forming the baseline the observation ablation must outperform."}],"review_version":1}