{"id":"e1f352d6-0b9b-42ca-bf99-9bf6d446eff1","arxiv_id":"2506.23768","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A new musculoskeletal dog model with 133 muscles, a centroid-based muscle line-of-action algorithm, and a differentiable muscle activation model achieves motion capture tracking with qualitative EMG agreement.","lead":"This paper builds a 133-muscle computer model of a dog in the MuJoCo physics engine and shows it can track real motion capture of walking, jumping, and sit-to-stand motions. It also smooths muscle activation dynamics to make gradient-based controllers converge, and compares the simulated muscle signals against published dog EMG recordings.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The EMG validation rests on per-muscle F0 values that are unreported and unvalidated; without a sensitivity analysis or physiological anchor, the claimed qualitative EMG agreement is not interpretable.","rationale":"The reader's weakest_assumption correctly identifies the unvalidated F0 calibration as the load-bearing vulnerability. I agree: the EMG validation is the paper's principal contribution, and its interpretability depends entirely on the muscle force capacities being biologically reasonable or at least not affecting the phase/relative patterns of excitation. The paper gives no evidence for either. The concern is not that the authors acted in bad faith; it is that a post-hoc calibration, unreported in detail, could easily be responsible for the observed agreement. My recommended verdict is unchanged from the reader's CONDITIONAL: the engineering contributions (procedural LoA extraction, sigmoid muscle dynamics, the model itself) are plausible and useful, but the scientific validation needs strengthening before the EMG claim can be accepted. A focused sensitivity analysis of F0 would settle whether this concern actually lands. I did not find a stronger concern: the kinematics tracking error is large but explicitly reported, the sigmoid dynamics claim is not ablated but is secondary to the EMG claim, and the lack of code/data release is a reproducibility issue that does not invalidate the internal reasoning.","tokens_in":13171,"tokens_out":3686,"duration_ms":45509,"concrete_test":"Run a sensitivity analysis on the walking clip: for each of the 10 muscles used in the EMG comparison, multiply its F0 by 0.5, 1.5, and 2.0 (one at a time, and also a global scaling sweep), re-run the iLQG tracking, and compare the phase of peak excitation and the set of muscles deemed 'active' versus the original. If the qualitative EMG conclusions (e.g., gastrocnemius active in stance, sartorius in swing) survive a factor-of-2 perturbation, the F0 concern is mitigated. If they flip, the claimed agreement is an artifact of the chosen scale. A complementary test is to recompute F0 from published canine PCSA data (e.g., Stark et al. 2021) and check whether tracking and EMG agreement still hold.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central scientific claim is that simulated muscle excitations 'align qualitatively' with canine EMG literature. This claim is read out directly from the control signals of an optimization that minimizes tracking error. The per-muscle peak force F0 (Eq. 2: F0 = scale / actuator_acc0) is a free parameter: the authors state that PCSA-based F0 estimates 'failed to accurately track the reference motion' and then switched to MuJoCo's built-in scaling, but they do not report the chosen scale values, the resulting F0 magnitudes, or any comparison to physiological force estimates. Because the iLQG controller can compensate for lower or higher force capacities by adjusting excitation, different F0 assignments will produce different excitation patterns even when kinematics track equally well. The EMG comparison is therefore uninterpretable without knowing that F0 is biologically plausible or that the excitation patterns are insensitive to F0. The paper also provides no error bars, no independent force validation, and no ablation isolating the sigmoid dynamics, but the F0 issue is the most direct threat to the EMG claim. A reviewer cannot distinguish physiological excitation from an artifact of the arbitrary scaling that happened to make tracking work.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces a parametric musculoskeletal dog model with 133 muscles generated procedurally from 3D muscle meshes, including a line-of-action extraction algorithm and a MuJoCo implementation. To make muscle activation/deactivation dynamics compatible with the iLQG planner, the authors propose a sigmoid-smoothed replacement for Millard's discontinuous switch (Eq. 5). They retarget 28 motion-capture clips of a Border Collie and use MJPC/iLQG to track walking, sprinting, sit-to-stand, and jumping motions. The central claims are that iLQG successfully tracks these clips with the proposed muscle dynamics, and that simulated muscle excitation patterns align qualitatively with previously published canine EMG data across the three tasks. The paper also acknowledges several mismatches (semimembranosus, vastus lateralis, flexor digitorum superficialis, and high biceps brachii activity during jumping) and plans to release the model and data.","tokens_in":13437,"tokens_out":5993,"duration_ms":63669,"significance":"The paper addresses an important integration of detailed anatomical modeling, differentiable control, and external EMG benchmarking. If the central claims are supported, the released model and task would be a useful community resource for quadruped biomechanics and neuromuscular control. The external EMG comparison is a genuine validation step rather than a circular fit to the simulation data, and the procedural line-of-action pipeline is a practical contribution. However, the current manuscript does not yet establish the two central claims: the per-muscle force scaling F0 is unreported and unvalidated, the sigmoid dynamics is not ablated, and the tracking and EMG comparisons lack quantitative criteria. These gaps are fixable, so the contribution is potentially strong but not yet proven.","major_comments":[{"comment":"The per-muscle peak force F0 is set as scale/actuator_acc0, but the manuscript does not report the chosen scale values, the resulting F0 magnitudes, or any comparison with physiological estimates. The PCSA-based alternative is dismissed only because \"the model failed to accurately track the reference motion,\" without reporting how far those values were from expected biology. Since the tracking optimization can compensate for force-capacity scaling by adjusting excitation, the excitation patterns used for EMG validation may be artifacts of this arbitrary scaling. The authors should report F0 for all muscles, compare it with PCSA times specific-tension estimates, and provide a sensitivity analysis (e.g., perturbing F0 by plus or minus 20 percent) demonstrating that the phase relationships reported for walking, sit-to-stand, and jumping are preserved. Without these data, the qualitative EMG claim is not interpretable.","section":"Methods, Model Creation (Eq. 2); Results, EMG comparison"},{"comment":"The Conclusions state that modifying the model to include differentiable muscle dynamics is \"essential\" for accurately tracking the motion capture clips, yet no experiment is reported comparing the proposed sigmoid-smoothed dynamics with the original Millard activation-deactivation dynamics under identical cost weights and planner settings. A head-to-head ablation should report tracking error, convergence behavior, and success or failure for the same clips with and without the sigmoid. This is a load-bearing claim for the paper's stated contribution of an improved muscle dynamics model.","section":"Conclusions; Methods, Eq. (5)"},{"comment":"The tracking performance is reported as average marker-position error over time, with maxima of about 0.4 m for walking, sprinting, and sit-to-stand and about 1 m for jumping. However, no error bars, number of repetitions, or clip-level statistics are given, and no baseline (e.g., kinematic retargeting without dynamics, a torque-actuated version of the same model, or chance-level error) is provided. To support the word \"successfully\" in the central claim, the authors should report per-clip distributions, standard deviations, and at least one baseline comparator, and should discuss whether a 1 m error for jumping is compatible with the subsequent EMG interpretation for that task.","section":"Results, Fig. 6"},{"comment":"The EMG validation is qualitative and the criteria for \"qualitative alignment\" are not defined. The text reports several clear mismatches: semimembranosus shows little activity, vastus lateralis is nearly constant, flexor digitorum superficialis lacks peaks during sit-to-stand, and biceps brachii is very active during jumping. The authors should provide a systematic comparison table listing, for each muscle and task, the expected phase of peak activation from the literature, the simulated peak phase, and a pre-specified agreement criterion (e.g., same gait sub-phase). This would allow readers to separate genuine agreement from selective reporting and to weigh the acknowledged exceptions.","section":"Results, EMG comparisons (walking, sit-to-stand, jumping)"}],"minor_comments":[{"comment":"Equation (5) uses \"sigmoid\" without defining the function or its argument convention; as written, at u = a the time constant is tau_d + (tau_a - tau_d) * sigmoid(0.5), which depends on the sigmoid's scale. Please define the sigmoid and state the intended value at the switch point.","section":"Methods, Eq. (5)"},{"comment":"The competing interests section contains the placeholder \"Related patent number here if applicable,\" which should be removed or completed.","section":"Competing interests"},{"comment":"The origin of the \"leo3Dmodels\" torque-actuated Pharaoh Hound model is mentioned but not cited; a reference or URL should be provided.","section":"Methods, Model Creation"},{"comment":"Figure 6 should include axis labels, units, and a legend identifying the plotted clips; the current figure does not allow the reader to reconstruct the reported error values.","section":"Results, Fig. 6"},{"comment":"In the sit-to-stand analysis, the time windows (frames 0-100, 100-300, etc.) are described only in text; annotating Figure 16 with these phases would make the comparison easier to evaluate.","section":"Results, Sit-to-stand"}],"recommendation":"major_revision","confidential_remarks":"This is a system-and-benchmark paper whose main future value is the released model and task. The two load-bearing gaps are the unvalidated F0 scaling and the missing sigmoid ablation; both are addressable within the manuscript's scope. I see no circularity concern: the EMG data are external to the optimization, and the self-citations to Barbera et al. and Demuth et al. provide methods context rather than the target result. The paper would be strengthened by a clearly defined EMG-agreement protocol before the qualitative claims can be assessed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I'll cut to the chase: this is a useful engineering contribution that overreaches one scientific claim. The new things are real: a 133-muscle MuJoCo dog built from 3D meshes, a simple centroid-slicing line-of-action extraction, a sigmoid-smoothed muscle activation dynamics, and a new Border Collie mocap set. The tracking results look plausible—within 0.4 m for walk/sprint/sit-to-stand, about 1 m for jump—and the model plus data will be a resource people actually use.\n\nThe soft spots are concentrated in the validation of muscle excitations against EMG. The reader is right that the per-muscle peak force F0 is the load-bearing modeling choice. The paper states PCSA-based F0 'failed to accurately track' and then switches to MuJoCo's built-in scale, but never reports the chosen scale or resulting F0 values. Since the optimization reads out excitations, and different F0 assignments can produce different excitation patterns even with equal tracking, the claimed 'qualitative agreement' with EMG is uninterpretable without either a physiological anchor for F0 or a sensitivity analysis. That's not a minor quibble; it's the difference between a modeling artifact and a biological finding.\n\nThe other issues are proportionately smaller but real. The claim that the sigmoid dynamics is 'essential' has no ablation—show me iLQG failing with the standard Millard switch. There are no error bars or baselines for the tracking errors. The EMG comparison is honest about its mismatches (semimembranosus, vastus lateralis, flexor digitorum superficialis), which is commendable, but the mismatches are tallied with no attempt to explain them. And the GitHub link is a plan, not a release; the paper says 'we plan to release,' so reviewers can't check the model.\n\nThe stress-test note holds up; I don't see a way around the F0 issue on the evidence given. The paper deserves a serious referee—the resource is valuable and the methods are clearly described—but the revision should report F0 values, add a sensitivity analysis or PCSA-benchmarked comparison, include an ablation for the sigmoid, and release the artifacts before the EMG claim is accepted.\n\nWho is this for? People building musculoskeletal models in MuJoCo, control researchers doing gradient-based motor control, and biomechanists who want a freely available canine platform. A reader in that group gets real value. My recommendation: send it to peer review, but expect heavy revision on the validation. I'd bring it to reading group as a good example of what a practical muscle model can do, with the caveat about F0.","headline":"Useful new dog model and muscle LoA pipeline, but the EMG validation rests on unmeasured muscle strengths (F0) and needs an ablation before the science holds up.","tokens_in":13966,"tokens_out":2243,"would_cite":false,"duration_ms":23066,"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":"A 133-muscle dog model built automatically from 3D muscle meshes tracks motion-capture clips and produces muscle activation timings that agree qualitatively with published canine EMG recordings.","keywords":["musculoskeletal model","canine locomotion","motion capture tracking","muscle dynamics","line of action extraction","model predictive control","iLQG","electromyography comparison"],"falsifier":"Repeat the walking clip with per-muscle peak forces recomputed from published canine PCSA and fibre-length values; if iLQG no longer tracks the motion, or if the gastrocnemius and triceps activation phases stop matching the EMG reports, then the reported EMG agreement depends on the unvalidated scaling rather than on the model's anatomy.","tokens_in":12971,"feed_emoji":"🐕","tokens_out":8622,"duration_ms":94210,"temperature":0.7,"pith_summary":"This paper sets out to show that a biomechanically plausible quadruped model can be built largely automatically from anatomical 3D muscle meshes, and that such a model can be driven by gradient-based predictive control to reproduce real dog movements. The authors construct a 133-muscle dog model by slicing each muscle mesh and connecting the slice centroids into a line of action, then add a motion-capture tracking task. They replace the standard muscle model's hard activation/deactivation switch with a smooth sigmoid, which makes the dynamics differentiable enough for the iLQG controller to track walking, sprinting, sit-to-stand, and jumping clips. The simulated muscle excitation patterns are then compared with published canine EMG data, with broad qualitative agreement for most of the ten muscles examined. If the claim holds, the work offers a reusable pipeline from anatomical geometry to controllable simulation for studying muscle function without new animal experiments.","feed_headline":"A 133-muscle virtual dog tracks motion capture and matches muscle signals","feed_subtitle":"Simulated muscle activations from an automatically built canine model agree with published recordings across gaits.","key_machinery":"The central machinery is a three-part pipeline. First, line-of-action extraction: each muscle mesh is sliced along its longitudinal axis at intervals proportional to muscle length; the centroid of each slice becomes a candidate via-point, kept when the path moves to a new bone or the gap since the last kept point exceeds a distance threshold, producing anatomically reasonable tendon routings without manual placement. Second, differentiable muscle dynamics: the Millard activation model's hard switch between activation and deactivation time constants is replaced by a sigmoid interpolation modulated by a smoothness parameter, so the activation dynamics are smooth enough for iLQG. Third, the motion-tracking task: a weighted cost of joint velocity, muscle excitation, marker-position error, and marker-velocity error is minimized by iLQG within the MuJoCo Predictive Control framework, converting raw motion-capture marker trajectories into muscle excitation signals.","core_discovery":"The paper's central claim is that a full musculoskeletal quadruped model does not have to be hand-tuned muscle by muscle: cutting each 3D muscle mesh into slices, connecting the slice centroids into a tendon path, and adding the resulting 133 actuators to a dog skeleton yields a model that a gradient-based predictive controller can drive through real motion-capture clips. The enabling modification is replacing the discontinuous Millard activation/deactivation switch with a sigmoid-smoothed transition, which gives iLQG a usable gradient. With this change, the controller tracks walking, sprinting, and sit-to-stand clips with marker errors within about 0.4 meters, and jumping with errors up to about 1 meter. The resulting excitation signals reproduce the phase patterns reported in canine EMG studies for most of the ten muscles compared, with the authors flagging specific exceptions such as vastus lateralis during walking, flexor digitorum superficialis during sit-to-stand, and notably high biceps brachii activity during jumping.","pith_inferences":["Because the per-muscle peak-force constants were taken from the simulator's built-in scaling rather than validated physiology, a sensitivity analysis sweeping those constants per muscle would reveal which of the reported EMG matches are stable and which depend on that scaling choice.","The validation is qualitative and marker-based, so the same model could be checked against force-plate ground reactions and joint moments; those data would test whether the kinematics are achieved with realistic kinetics, not just realistic-looking paths.","The sigmoid-smoothed muscle dynamics are not specific to dogs: applying the same smoothing to other MuJoCo muscle models could make iLQG practical for other high-dimensional musculoskeletal systems and could also ease gradient-based reinforcement learning.","If the F0 scaling can be tied to published physiological cross-sectional areas and fibre lengths, the tracking setup could be inverted into a tool for estimating muscle forces during locomotion, not just activation timings."],"forward_implications":["The same pipeline can be rerun from fresh muscle meshes to produce a controllable musculoskeletal model without manual muscle-path authoring, so other dog breeds or other quadrupeds become practical to model.","The smooth activation dynamics remove the main gradient barrier that kept iLQG from controlling 133-muscle models, making model-based predictive control a viable alternative to reinforcement learning for musculoskeletal tracking.","Simulated excitation timings can serve as testable predictions for muscles and gaits where experimental EMG is missing, and the planned release of the model and 28 motion-capture clips lets other groups reproduce or extend those predictions.","Tracking marker positions within about 0.4 meters for walking, sprinting, and sit-to-stand suggests the model can be used to study muscular demands of everyday behaviours, while the larger jumping error marks jumping as a stress test requiring better force or contact modelling."],"supporting_citations":[{"why":"Supplies the MuJoCo physics engine in which the model and its muscle actuators are simulated.","marker":"Todorov et al. (2012)"},{"why":"Defines the activation/deactivation muscle dynamics that the paper replaces with a sigmoid-smoothed version.","marker":"Millard et al. (2013)"},{"why":"Provides the iLQG trajectory optimizer used to solve the motion-tracking task.","marker":"Tassa et al. (2012)"},{"why":"Provides the MuJoCo Predictive Control framework through which iLQG is applied.","marker":"Howell et al. (2022)"},{"why":"Supplies the canine limb EMG recordings used as the qualitative comparison target for simulated excitations.","marker":"Goslow et al. (1981)"},{"why":"Supplies EMG activity patterns for extrinsic limb muscles across gaits, compared for sartorius and biceps femoris.","marker":"Deban et al. (2012)"},{"why":"Supplies sit-to-stand muscle activation data used to validate the sit-to-stand simulation.","marker":"Ellis et al. (2018)"},{"why":"Presents the prior volumetric line-of-action method that the paper's centroid-slicing algorithm is designed to replace.","marker":"Demuth et al. (2022)"},{"why":"Presents a prior dog musculoskeletal model validated against EMG, giving a comparative benchmark for the new model.","marker":"Stark et al. (2021)"},{"why":"Presents the earlier ostrich mocap-tracking musculoskeletal task whose design the dog task follows.","marker":"Barbera et al. (2021)"}],"fun_headline_variants":["Virtual dog with 133 muscles tracks motion, matches EMG","Automatically built canine model reproduces EMG patterns","Smoothed muscle switch lets iLQG steer dog model","133-muscle dog model nails motion capture and EMG"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the maximum force assigned to each muscle, which was taken from the simulator's built-in scaling rather than from measured dog physiology, is close enough to reality that the simulated activation timings can meaningfully be compared with real muscle recordings.","fun_headline_variants_meta":{"raw":{"variants":["Virtual dog with 133 muscles tracks motion, matches EMG","Automatically built canine model reproduces EMG patterns","Smoothed muscle switch lets iLQG steer dog model","133-muscle dog model nails motion capture and EMG"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000138,"raw_usage":{"total_tokens":1113,"prompt_tokens":862,"completion_tokens":251,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":478,"completion_tokens_details":{"reasoning_tokens":182}},"tokens_in":478,"tokens_out":251,"duration_ms":3111,"temperature":1.0,"reasoning_tokens":182,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T21:31:49.758213+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Repeat the walking clip with per-muscle peak forces recomputed from published canine PCSA and fibre-length values; if iLQG no longer tracks the motion, or if the gastrocnemius and triceps activation phases stop matching the EMG reports, then the reported EMG agreement depends on the unvalidated scaling rather than on the model's anatomy.","supporting_citations":[],"review_version":1}