{"id":"b8d0226e-acd1-46f1-b41f-356b90047cdd","arxiv_id":"2505.13436","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A single imitation-learning policy tracks the movements of 467 people, including people with gait impairments, with torque- and muscle-driven biomechanical models and estimates joint torques, ground reaction forces, and muscle activations from markerless video.","lead":"KinTwin trains a reinforcement learning policy to make a biomechanical human body model in a physics simulator copy movements captured on video, then reports the forces, torques, and muscle activations the copy requires. Clinicians could use such a system to get rich movement measurements from ordinary cameras instead of instrumented labs.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reward sign error and unreported residual-force magnitudes undermine the unvalidated kinetic claims, so the conditional verdict stands","rationale":"The paper is strongest on the kinematic side: the policy is evaluated on held-out participants, reports joint-angle errors below 1 degree, and the instrumented-walkway comparison (Sec. 4.2) checks event timing and stride length against an independent floor sensor. Those results support the 'accurate replication' part of the claim. The kinetic part is where the central claim extends beyond kinematics, and it is exactly there that the support is thinnest. The residual force control is introduced as a way to handle external forces (Sec. 3.3), but the reward equation (5) as printed rewards rather than penalizes large residual forces, and no ablation or diagnostic reports the scale of these forces. If the policy leans on the residual force, then the joint torques and muscle activations shown in Figures 1 and 3 are outputs of the non-physical actuator, not estimates of the subject's internal kinetics. The instrumented walkway validation cannot rule this out because it only compares timing and stride length of thresholded vertical GRF, not force magnitudes or joint moments. The paper itself, in Sec. 5.1, states that substantial work remains to tune the torques and muscle activations and that 3D GRF validation against force plates is a future direction, which is a fair caveat, but it also means the abstract's 'clinically meaningful differences' are currently unverified. This is a testable concern, not a reason to reject: quantifying the RFC forces and adding a ground-truth kinetic comparison (force plates, EMG) would settle it. The sign error in Eq. (5) is a concrete, checkable internal inconsistency that strengthens the concern. Therefore the CONDITIONAL verdict is appropriate; no change is needed.","tokens_in":17189,"tokens_out":4266,"duration_ms":38771,"concrete_test":"On the existing 397-trial test set, compute per-rollout the norm of the first six action components ‖a_t^(0:6)‖ (RFC force) at each step, as a fraction of body weight, split by trials with/without assistive devices or therapist contact. If RFC exceeds ~5% BW during unassisted walking, or correlates with inferred hip/knee torques, the non-physical force is doing mechanical work and the kinetic differences are not from muscles. Also re-run with the sign in Eq. (5) corrected to +w_RFC‖a‖² and compare torque outputs; material changes would show the reported kinetics depend on the sign error.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The kinetic claim ('clinically meaningful differences in joint torques and muscle activations') depends on the residual force control (RFC) of Sec. 3.3 acting as a proxy for external forces and not contributing to lower-limb joint kinetics. Two problems. First, Eq. (5) has a sign inconsistency: the text says RFC is 'discourag[ed]', but the term is −w_RFC‖a_t^(0:6)‖², which rewards large residual forces; with w_RFC=0.075 this is an active incentive, not a penalty. Second, no experiment reports the magnitude of the 6-DoF pelvis residual force, and no comparison against force-plate/EMG ground truth is offered; the instrumented-walkway validation in Sec. 4.2 only checks event timing and stride length from thresholded vertical GRF, not force magnitudes or joint moments. The paper's own Limitations (Sec. 5.1) concede that 'substantial work will be required to fully tune both the torques and muscle activations' and that 3D GRF validation is 'an important future direction.' As written, the policy can exploit a large non-physical pelvis force to track the target trajectory, and the inferred torques/activations in Figs. 1 and 3 would then reflect the residual actuator rather than the subject's muscles. This is the load-bearing weakness in the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents KinTwin, an imitation learning system that tracks human movement in a physics simulator using a biomechanical human model. The system is trained with PPO on a large dataset of markerless motion capture from 467 participants, including many with movement impairments, and is evaluated on 41 held-out participants. The authors use both a torque-driven model and a muscle-driven model with 92 lower-limb muscles, and they introduce residual force control (RFC) at the pelvis to handle external forces such as canes, walkers, or therapist assistance. Kinematic tracking errors are small (0.65 deg mean joint angle error for the torque model) and gait-event timing errors against an instrumented walkway are moderate (80 ms foot contact, 20 mm stride length). The paper further claims that the inferred joint torques and muscle activations capture clinically meaningful differences, with qualitative examples from a transfemoral amputee and a person with hemiparetic gait. The central claim is that imitation learning can provide a 'digital kinetic twin' from markerless video, making full kinetic analysis broadly accessible.","tokens_in":17524,"tokens_out":4221,"duration_ms":41418,"significance":"If the kinetic inferences were valid, this would be an important contribution to clinical movement analysis, as it would allow joint torques and muscle activations to be estimated from widely available video. The kinematic results are solid: the held-out participant split, the large dataset, and the instrumented-walkway comparison are clear strengths, and the 0.65 deg angular error and 20 mm stride error are impressive. The use of a muscle-driven model with GPU acceleration is a technical advance. However, the kinetic claims are currently supported only by qualitative examples and by a residual-force mechanism whose effect on the estimated kinetics is not quantified. The paper's own limitations section acknowledges that substantial tuning and 3D GRF validation remain. The significance is therefore conditional on resolving the residual-force issue and either validating the kinetics or scaling back the conclusions.","major_comments":[{"comment":"Eq. (5) contains a sign error that is load-bearing for the kinetic interpretation. The text states that the term '-w_RFC ||a_t^(0:6)||^2' is an 'additional penalization to discourage using the RFC actions', but a negative squared norm rewards large residual forces: with w_RFC=0.075, minimizing la with respect to the RFC action increases the magnitude of the 6-DOF pelvis force. The policy is therefore actively incentivized to apply non-physical assistance, rather than to minimize it. The paper never reports the magnitude of the residual force for any trial, so the reader cannot determine whether the inferred joint torques and muscle activations in Figures 1 and 3 reflect the subject's neuromuscular output or the residual actuator. Please correct the sign (if a penalty is intended), report the distribution of residual forces in the test set, and relate the magnitudes to plausible external forces (e.g., cane or therapist loads).","section":"Section 3.5, Eq. (5)"},{"comment":"The abstract and Section 5 state that the policy can infer 'clinically meaningful differences in joint torques and muscle activations', but the experimental evidence does not support this claim. The instrumented-walkway validation in Section 4.2 only checks the timing of thresholded vertical GRF events and stride length; it does not validate force magnitudes, mediolateral or anteroposterior GRF, joint moments, or muscle activations. The Limitations section (5.1) concedes that 'substantial work will be required to fully tune both the torques and muscle activations' and that validating 3D GRFs against force plates is 'an important future direction'. Without force-plate or EMG ground truth, the 'clinically meaningful differences' in Figures 1 and 3 are qualitative and could arise from the policy's non-physical residual forces. Please either add validation on a subset of trials or temper the abstract and conclusions to state that the kinetic outputs are plausible but unvalidated.","section":"Sections 4.2, 5.1, and abstract"},{"comment":"The reported tracking accuracy is partly conditioned on the fact that the target trajectories are produced with the same biomechanical model used in the simulator. Section 7.2 states: 'The biomechanical model used to reconstruct their kinematic trajectories was the same one used in the imitation learning environment.' Consequently, the target and the simulator share the same kinematic chain, joint limits, and body scaling, so the 0.65 deg joint-angle error is not a fully independent measure of how well KinTwin reproduces markerless motion capture; it measures how well the policy tracks trajectories that were themselves fitted with the identical model. The instrumented-walkway comparison is an external anchor and is genuinely valuable, but the paper should acknowledge this circularity explicitly and, where possible, report the markerless fitting error (e.g., reprojection error) so readers can estimate the component of tracking error attributable to the motion-capture pipeline.","section":"Section 7.2, Table 1"}],"minor_comments":[{"comment":"The phrase 'our work differences in several ways' should be 'our work differs in several ways'.","section":"Abstract"},{"comment":"The word 'faluers' should be 'failures'.","section":"Table 1 caption"},{"comment":"The phrase 'the different between the stride lengths' should be 'the difference between the stride lengths'.","section":"Section 7.4"},{"comment":"The list of body scale parameters reads 'an overall size, the pelvis, left thigh, left leg and foot, right thigh, right leg and foot, the left arm, and the left leg'; the final 'left leg' appears to be a typo, likely intended to be 'right arm' or another segment. Please check.","section":"Section 7.1"},{"comment":"Joint torques are reported in arbitrary units (a.u.) without any normalization or conversion factor. This limits clinical interpretability; please state how the torques are scaled or provide a calibration to SI units if one exists.","section":"Figures 1 and 3 and Table 1"},{"comment":"Metrics are reported as point estimates without measures of variability. Given 41 test participants, reporting standard deviations or bootstrap confidence intervals would substantially strengthen the evaluation.","section":"Tables 1 and 2"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is technically interesting and the kinematic results are strong, but the kinetic claims go well beyond the present evidence. The sign error in Eq. (5) is a concrete flaw that must be fixed or carefully justified, and the residual-force magnitudes need to be reported. If the authors can correct the reward formulation, report residual forces, and either add kinetic validation or substantially temper the abstract and conclusions, the paper could become acceptable. The lack of a public dataset or code also limits reproducibility, though that is not itself a reason for rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is worth a serious look, but read the kinetics section with suspicion. The kinematic core is solid. KinTwin trains one whole-body policy on torque- and muscle-driven biomechanical models from markerless motion capture of 467 participants, including many with impaired mobility, and evaluates on 41 held-out participants. The torque model tracks 0.65 deg joint angle error, foot-contact timing error of 80 ms, and stride length error of 20 mm against an instrumented walkway. Those numbers are good, and the lesion studies (removing future observations or RFC) make the contribution believable. The muscle-driven model is less accurate (1.68 deg, 120-135 ms) but still a reasonable first demonstration. The paper is honest about many of its own limitations.\n\nThe problem is the central kinetic claim. The abstract says the policy infers clinically meaningful differences in joint torques and muscle activations, but that is not backed by any force-plate or EMG ground truth. The 6-DoF residual force applied to the pelvis is never reported in magnitude, and the walkway validation only checks event timing and stride length from thresholded vertical GRF. On top of that, Eq. (5) has a sign error: the text says the RFC action is penalized, but the term is -w_RFC||a^(0:6)||^2, which rewards large residual forces. With w_RFC = 0.075, that is not a small incentive. As written, the policy could be using the non-physical pelvis force to make up for anything the legs cannot generate, so the \"clinically meaningful\" torque and activation differences in Figures 1 and 3 could be an artifact of the residual actuator. The paper's own limitations section concedes that substantial tuning is needed and that 3D GRF validation is future work, but the abstract overstates what is shown.\n\nThe circularity concern is real but moderate: the target trajectories are fit with the same biomechanical model used in the simulator, so the low tracking error partly reflects self-consistency. The instrumented walkway comparison provides an external anchor for timing and stride, but not for forces.\n\nMy recommendation: send this to peer review, but the reviewers should press for a corrected Eq. (5), a report of residual-force magnitudes, and a clearly labeled separation between demonstrated kinematics and unvalidated kinetics. The kinematic contribution alone is substantial enough to justify publication; the kinetic claims need to be toned down or validated. I would not want this paper to be rejected on the kinematic results because of the overreach in the abstract, and I would not want it accepted without fixing the kinetic overreach.","headline":"Kinematic replication is solid and worth peer review; the clinically meaningful kinetic claims are not yet supported and need a corrected Eq. (5), residual-force reporting, and force-plate/EMG validation.","tokens_in":17987,"tokens_out":3340,"would_cite":true,"duration_ms":31222,"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":"KinTwin shows that one imitation-learning policy can turn markerless video of able-bodied and impaired movement into accurate kinematics, joint torques, and muscle activations.","keywords":["imitation learning","biomechanical model","markerless motion capture","inverse dynamics","muscle-driven simulation","gait analysis","rehabilitation","digital twin"],"falsifier":"Run KinTwin on a cohort of walkers captured simultaneously with markerless video, force plates, and surface EMG, and compare the inferred joint torques and muscle activations against the measured values; if the inferred peak hip or knee moments differ by more than a standard gait laboratory's measurement error, or the muscle-activation timing diverges systematically, the claim of clinically meaningful kinetic inference is falsified. A simpler preliminary check is to inspect the pelvis residual force in trials with no assistive device, since large forces there would show the policy depends on non-physical assistance.","tokens_in":16983,"feed_emoji":"🎥","tokens_out":9247,"duration_ms":79484,"temperature":0.7,"pith_summary":"This paper claims that a single imitation-learning policy, trained on a large biomechanical dataset of able-bodied and impaired movement, can solve inverse dynamics from markerless motion capture: given only video-derived joint trajectories, it produces the joint torques and muscle activations that would generate them. The authors train the policy on 34 hours of data from 467 participants, most with movement impairments such as amputations or neurological conditions, and show it replicates unseen participants' kinematics with about a degree of average joint-angle error. They also validate the inferred ground reaction forces against an instrumented walkway, matching foot-contact and foot-off timing to within about 80 ms and stride length to within about 20 mm. If correct, this would make kinetic gait analysis, normally requiring force plates and motion-capture laboratories, available from ordinary multi-camera video, enabling detailed characterization of impairment and response to treatment.","feed_headline":"KinTwin pulls joint torques and muscle activity from plain video","feed_subtitle":"A single policy replicates able-bodied and impaired motion from markerless video, matching gait events to about 80 ms.","key_machinery":"The load-bearing mechanism is a goal-conditioned imitation-learning policy that maps the simulator state and future samples of the target trajectory, at look-ahead times of 0, 1, 2, 4, 8, and 16 steps, to either joint torques or muscle activations. Two named ingredients carry the argument. First, residual force control (RFC) adds a non-physical six-degree-of-freedom force to the pelvis, letting the policy track movements that involve canes, walkers, therapist hands, or chairs without abandoning the target trajectory; an ablation without RFC raises the failure rate from 3.8 percent to 61.2 percent. Second, the biomechanical models are individualized to each participant via eight body-scale parameters, which is what lets one policy replicate many different body shapes and impairments. Training uses dense tracking rewards, squared-error pose and velocity losses, plus action penalties that discourage large or fast-changing residual forces.","core_discovery":"The central discovery is that a single goal-conditioned imitation-learning policy can closely track the kinematics of both able-bodied and impaired movement in a physics simulator and, in doing so, infer the kinetics behind the motion. Using a torque-driven biomechanical model, the policy tracks unseen test participants with mean joint-angle errors below one degree and horizontal pelvis errors near 4 cm across a range of mobility assessments; a muscle-driven model with 92 lower-limb muscles tracks with 1.68 degrees mean joint-angle error. The inferred vertical ground reaction forces match instrumented-walkway events with median foot-contact errors of about 80 ms and stride-length errors of about 20 mm. The paper further reports that the inferred torques and muscle activations differ across clinical conditions in expected ways, such as reduced propulsion on the prosthetic side of a transfemoral amputee and reduced hamstring activation on the hemiparetic side after stroke, indicating sensitivity to clinically meaningful features.","pith_inferences":["The residual force's magnitude and direction, currently a modeling artifact, could be reinterpreted as an estimate of external support load from canes, walkers, or therapist assistance, turning KinTwin into a tool for quantifying assistive-device use during clinical assessments.","Since the policy must use larger residual forces to track poorly balanced or unstable movements, the residual-force signal might serve as a video-derived proxy for fall risk or balance impairment, a hypothesis the paper does not test.","Combining the 30 Hz markerless trajectories with higher-rate inertial or video data could sharpen stance-phase and foot-off timing estimates, addressing the temporal-resolution limitation the paper acknowledges.","The same training pipeline could be applied to longitudinal data from individual patients to detect changes in torque or activation asymmetry before they become visible in kinematics, an extension the paper frames only as future characterization work."],"forward_implications":["If correct, the approach lets clinicians obtain joint torques, ground reaction forces, and muscle activations from multi-camera video, without instrumented walkways or force plates.","The reported tracking accuracy suggests the policy could quantify spatiotemporal gait parameters, such as stride length and foot-contact timing, across a population that includes amputees and people with neurological impairments.","The muscle-driven variant shows that asymmetric activation patterns, such as reduced hamstring drive on a hemiparetic side, can be extracted from video, opening a path to video-based muscle assessment without wearable sensors.","Because the policy is validated on held-out participants, the method may generalize to new patients with similar impairments, supporting longitudinal tracking of recovery or intervention response."],"supporting_citations":[{"why":"Supplies the differentiable biomechanical reconstruction approach used to obtain kinematic trajectories from markerless video.","marker":"[3]"},{"why":"Provides the GPU-accelerated physics engine that made training at four to eight billion simulation steps feasible.","marker":"[11]"},{"why":"Defines the torque- and muscle-driven biomechanical models, including the 92 lower-limb muscles, used in the imitation-learning environment.","marker":"[13]"},{"why":"Introduces residual force control, the mechanism this paper relies on to track trajectories involving external support forces.","marker":"[20]"},{"why":"Demonstrates that physics-based motion imitation improves spatiotemporal gait parameters from video, the baseline this work extends to biomechanical models and impaired populations.","marker":"[24]"},{"why":"Provides the policy-optimization algorithm used to train the imitation policy.","marker":"[25]"},{"why":"Supplies the reinforcement-learning environment and implementation used for large-scale parallel training.","marker":"[26]"}],"fun_headline_variants":["KinTwin: from markerless video to precise joint torques and muscle forces","Imitation learning turns video into biomechanical analysis for impaired gaits","KinTwin matches gait events to 80 ms from plain video, inferring muscle activity","Single policy replicates able-bodied and impaired motion, extracting kinetics from video","KinTwin infers muscle forces and joint torques from plain video"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The kinetic outputs are clinically meaningful only if the invisible extra force the policy applies to the pelvis truly stands in for external support like canes, walkers, or therapist hands, and does not quietly do work that should come from the legs; the paper does not report how large that force is for any trial or compare inferred torques or muscle activations against force-plate or EMG measurements.","fun_headline_variants_meta":{"raw":{"variants":["KinTwin: from markerless video to precise joint torques and muscle forces","Imitation learning turns video into biomechanical analysis for impaired gaits","KinTwin matches gait events to 80 ms from plain video, inferring muscle activity","Single policy replicates able-bodied and impaired motion, extracting kinetics from video","KinTwin infers muscle forces and joint torques from plain video"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000687,"raw_usage":{"total_tokens":3148,"prompt_tokens":1014,"completion_tokens":2134,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":630,"completion_tokens_details":{"reasoning_tokens":2034}},"tokens_in":630,"tokens_out":2134,"duration_ms":15652,"temperature":1.0,"reasoning_tokens":2034,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:13:51.620584+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run KinTwin on a cohort of walkers captured simultaneously with markerless video, force plates, and surface EMG, and compare the inferred joint torques and muscle activations against the measured values; if the inferred peak hip or knee moments differ by more than a standard gait laboratory's measurement error, or the muscle-activation timing diverges systematically, the claim of clinically meaningful kinetic inference is falsified. A simpler preliminary check is to inspect the pelvis residual force in trials with no assistive device, since large forces there would show the policy depends on non-physical assistance.","supporting_citations":[{"cited_title":"Residual Force Control for Agile Human Behavior Imitation and Extended Motion Synthesis","cited_arxiv_id":null,"evidence_quote":"Introduces residual force control, the mechanism this paper relies on to track trajectories involving external support forces."}],"review_version":1}