PaperlayerView exact revision on arXiv

Source-linked AI summary

A generalist policy for muscle-driven control across individual musculoskeletal models

Bozhao Wang, Yinglei Zhu, Zhengqing Li, Zhiyuan Zhang, Weiming Kang, Yiheng He, XinJun Liu, Senhao Hou, Huichan Zhao, Xiaoqiang Tang

arXiv:2610.36535v1cs.RO

TL;DR

Differences in musculoskeletal morphology raise the question of whether a muscle-driven policy can track reference motions across individual models. MyoAcross combines individual model construction, morphology-conditioned policy learning and virtual model augmentation; it achieves 99.24% completion versus 98.24% for an unconditioned MLP, with additional tests on biomechanical measurements and model variation.

  • Problem

    Differences in skeletal proportions, muscle geometry and mass change the dynamics controllers encounter, leaving open whether policies can transfer across individual models excluded from training.

    Source: Introduction · p. 2

  • Method

    MyoAcross constructs individual musculoskeletal models and trains a generalist muscle-driven policy conditioned on model-derived morphology features, with virtual model augmentation.

    Source: Introduction · p. 2, Individual musculoskeletal models… · p. 4

  • Results

    MyoAcross reached 99.24% completion versus 98.24% for the unconditioned MLP, while comparisons found agreement in vertical ground reaction force waveforms and channel-dependent agreement in muscle activity.

    Source: A generalist policy enables reuse… · p. 4, Discussion · p. 11

  • Takeaways & Limitations

    The common policy interface supports comparative musculoskeletal simulations across individual models while holding policy parameters and source-motion selection fixed.

    Source: Discussion · p. 11

Abstract

from arXiv · show

Differences in musculoskeletal morphology challenge policy reuse across individual models. Here we present MyoAcross, a framework that combines individual musculoskeletal model construction, morphology-conditioned policy learning and virtual model augmentation to enable muscle-driven reference-motion tracking without individual fine-tuning. Across 31 held-out individual models, MyoAcross achieves 99.24% completion and reduces tracking error by 22.60% relative to an unconditioned multilayer perceptron. Policy reuse extends to models of 22 held-out individuals constructed from single-image shape estimates, measured height and body mass, and, with asymmetry augmentation, to models with unilateral lower-limb shortening. Comparisons with measurements from 120 additional individuals show agreement in vertical ground reaction force waveforms and channel-dependent agreement between simulated muscle activations and electromyography. These results establish a common control framework for comparative simulations of movement across individual musculoskeletal models.

Introduction

Differences in individual musculoskeletal morphology raise the question of whether muscle-driven policies can track reference motions across models excluded from training. MyoAcross combines individual model construction, morphology-conditioned control and virtual model augmentation to examine that question, including transfer to asymmetric and image-derived models.

Introduction

  • Differences in skeletal proportions, muscle geometry and mass change the dynamics encountered by controllers, motivating tests of policy transfer across individual models.Source: Introduction · p. 2
  • The study asks whether morphology conditioning supports muscle-driven tracking of time-resolved reference motions across individual models, including individuals excluded from training.Source: Introduction · p. 2
  • Unilateral lower-limb shortening tests policy reuse under structural variation while retaining the original reference motions; asymmetry augmentation is also examined as a transfer factor.Source: Introduction · p. 2
  • MyoAcross combines individual musculoskeletal model construction with generalist muscle-driven policy learning, conditioning the policy on morphology features and using virtual model augmentation.Source: Introduction · p. 2
  • Further tests assess sensitivity to morphology-feature and reference-motion input errors and reuse on models built from single-image shape estimates, measured height and body mass.Source: Introduction · p. 3

Results

MyoAcross combines individualized musculoskeletal models with morphology-conditioned muscle control and virtual model augmentation. Results test reuse across held-out, asymmetric, image-derived, and measured-subject models, alongside biomechanical agreement and input sensitivity.

Individual musculoskeletal models support a common policy interface

  • Regional scaling fitted 17 skeletal measurements with 2.51 mm mean absolute error across 141 models, versus 13.50 mm for global scaling and 24.97 mm for the unchanged template.Models retained common topology and a 416-dimensional muscle excitation space; morphology features conditioned the policy through FiLM.Source: Individual musculoskeletal models… · p. 4

A generalist policy enables reuse across held-out individual models

  • MyoAcross reached 99.24% completion and 53.70 mm tracking error, reducing error 22.60% versus the unconditioned MLP on 31 held-out models without fine-tuning.All three multi-model policies shared the training model bank, reference motions, and interaction budget.Source: A generalist policy enables reuse… · p. 4, Figure 3 · p. 5
  • MyoAcross had lower tracking error than both multi-model baselines on all 31 held-out individuals and outperformed the unconditioned MLP in all eight motion categories.It outperformed Concat in six categories; Concat had lower error in slow walking and right turns.Source: A generalist policy enables reuse… · p. 4, Figure 3 · p. 5

Biomechanical assessment of walking across individual models

  • Across 101 subjects, MyoAcross’s simulated vertical ground reaction forces had mean correlation 0.853 and body-weight-normalized RMSE 0.139 against measurements.Muscle-activation agreement with EMG was channel-dependent: mean correlation was 0.406 across 12 channels, with soleus channels highest and some channels negative.Source: Biomechanical assessment of walking… · p. 6, Figure 4 · p. 7

Virtual model augmentation improves generalization to held-out individual models

  • With the same interaction budget, virtual model augmentation raised completion from 89.21% to 99.24% and reduced tracking error from 76.99 to 53.70 mm on held-out models.Completion increased in 30 of 31 individuals and tracking error fell in all 31.Source: Virtual model augmentation improves… · p. 6, Figure 5 · p. 8

Asymmetry augmentation extends policy reuse to asymmetric test models

  • At 10% unilateral shortening, asymmetry augmentation raised completion from 1.56% to 91.78% on the left and from 20.51% to 88.41% on the right.On original models, both policies achieved 99.24% completion, but tracking error was higher with asymmetry augmentation: 71.20 versus 53.74 mm.Source: Asymmetry augmentation extends policy… · p. 6, Figure 5 · p. 8

Policy sensitivity to morphology feature and reference-motion input errors

  • MyoAcross completion was 97.85% under 5% morphology-feature bias, but fell to 70.51% under a 5° reference-motion offset and 18.57% under combined 10% and 10° perturbations.At the largest combined perturbation, reference coverage was 58.68%.Source: Policy sensitivity to morphology… · p. 9

Single-image model construction provides an entry point for policy reuse

  • Across 88 image-derived models, MyoAcross completed 97.90% of reference motions, with 99.34% reference coverage and 54.38 mm tracking error.For the 22 individuals in the original held-out split, completion was 99.29%.Source: Single-image model construction… · p. 9

Discussion

MyoAcross provides a shared controller for comparative tracking across individual musculoskeletal models, with biomechanical comparisons and tests of prescribed asymmetry. Its demonstrated scope is bounded by shared model topology and unresolved fidelity of image-derived models.

Discussion

  • MyoAcross supports muscle-driven reference-motion tracking across held-out individual models without individual fine-tuning, providing a common controller for comparative musculoskeletal simulations.The paper reports that morphology conditioning and virtual model augmentation improve tracking across these models under a common interaction budget.Source: Discussion · p. 9
  • Morphology conditioning supplies model-specific information, while virtual model augmentation exposes the policy to broader dynamics; together they support excitations conditioned on morphology and musculoskeletal state.Geometry, mass distribution, inertia and musculotendon mechanics can make similar states require different excitations.Source: Discussion · pp. 9, 11
  • Errors in morphology features and reference-motion inputs can impair tracking even when the physical model is unchanged.The sensitivity tests varied policy inputs while keeping the physical model fixed.Source: Figure 6 · p. 10, Discussion · p. 11
  • Gait120 comparisons show agreement in vertical ground-reaction-force waveforms and channel-dependent agreement in muscle-activity patterns; asymmetry augmentation extends control to prescribed unilateral lower-limb shortening.The shortening tests were motivated by clinical reports of limb-length differences [20,21].Source: Discussion · p. 11
  • Demonstrated generalization is limited to models with a common topology and source-motion selection; individual muscle strength, tissue properties and neural control require further measurements or modelling assumptions.The asymmetric tests use prescribed unilateral shortening rather than the full range of neuromuscular changes associated with clinical conditions.Source: Discussion · pp. 11–12
  • Image-derived models’ anatomical and physiological fidelity remains to be assessed against individual measurements; further work could examine individually measured muscle function under matched movement conditions.The authors describe MyoAcross as a foundation for such studies.Source: Discussion · p. 12

Methods

The study reanalysed publicly available datasets under their respective licence terms, relying on approvals and consent from the original studies. No new participants or recordings were collected.

Ethics

  • The analysis used publicly available body-shape, motion, video and physiological data from AMASS, MoVi and Gait120.Data were accessed under their respective licence terms.Source: Ethics · p. 12
  • The original MoVi and Gait120 studies had ethics approval, and participants provided written informed consent.Gait120 consent included permission for data publication.Source: Ethics · p. 12
  • The study recruited no new participants and collected no new motion-capture or physiological recordings; additional ethical approval was not required for this secondary analysis.Source: Ethics · p. 12

Control formulation

The control formulation combines muscle activation with multibody dynamics and represents each individual model through physical parameters and state. A morphology-conditioned policy can adapt its excitation distribution to the model while sharing policy parameters.

Control formulation

  • Each model’s geometry, mass distribution, inertia and musculotendon properties shape coupled muscle and multibody dynamics, through which muscle excitations produce movement.Muscle tension depends on activation, muscle length and velocity.Source: Control formulation · p. 12
  • Model-derived morphology features summarize individual physical parameters and are added as a separate conditioning input to the policy.The policy observations include musculoskeletal states, reference-motion targets and derived tracking information.Source: Control formulation · p. 12
  • With identical observations, an unconditioned policy produces the same excitation distribution across models, whereas conditioning lets that distribution depend on the individual model while sharing parameters.This addresses cases where model differences require different excitations to track the same target.Source: Control formulation · p. 12

Individual musculoskeletal model construction

Individual musculoskeletal models are built by fitting skeletal proportions to shape-derived measurements and calibrating physical properties to body mass. The construction retains a common template topology while adapting regional geometry and associated parameters.

Individual musculoskeletal model construction

  • The construction fits 13 regional scale factors to 17 skeletal measurements derived from SKEL, adapting proportions while preserving template connectivity and muscle–tendon organization.A single global scale factor fitted to the same measurements serves as a uniform-scaling comparator.Source: Individual musculoskeletal model… · p. 12
  • Under density-preserving geometric similarity, regional dimensions update segment volume, mass distribution and inertia, while total mass is calibrated to recorded body mass.Musculotendon operating lengths follow adapted paths, and muscle force capacity scales with physiological cross-sectional area under constant specific tension.Source: Individual musculoskeletal model… · p. 13
  • For Gait120, SMPL-H shape parameters came from labelled marker recordings; for the single-image experiment, HSMR estimated SKEL shape from one photograph, calibrated to measured height and combined with body mass.Both routes use target skeletal measurements and the same individual-model construction procedure.Source: Individual musculoskeletal model… · p. 13

Subjects and reference motions

The study split 141 AMASS subjects by identity and used shared KIT source motions to build individual reference trajectories. A separate 120-subject dataset supplied biomechanical assessment data.

Subjects and reference motions

  • 141 AMASS subjects were split into 110 training and 31 evaluation identities, independently of the shared source-motion collection.Each subject’s body shape and recorded mass were used to construct an original individual musculoskeletal model.Source: Subjects and reference motions · p. 13
  • 64 KIT source motions covered eight categories, with eight clips per category, and were retargeted to each original model using GMR-Fit.GMR-Fit combines General Motion Retargeting with SMPL fitting to generate time-resolved reference trajectories.Source: Subjects and reference motions · p. 13
  • Biomechanical assessment used 960 retargeted medium-speed walking references for 120 Gait120 subjects, whose recorded level-walking GRF and EMG data were used for comparison.The common source-motion selection was not matched to each subject’s recorded walking speed or cadence.Source: Subjects and reference motions · p. 13

Morphology-conditioned policy

The policy conditions its actor and critic on each model’s morphology through FiLM, while comparison policies test conditioning and feature-integration choices.

Morphology-conditioned policy

  • Each compiled model was represented by 17 skeletal measurements and total mass, log-transformed and normalized using fixed statistics from candidate training models.Original, virtual, and asymmetric test models each supplied their own morphology features.Source: Morphology-conditioned policy · p. 13
  • A morphology encoder produced a learned representation that FiLM used to scale and shift normalized hidden features in the actor’s and critic’s separate residual backbones.Observations included musculoskeletal states, motion progress, current and future targets, and derived tracking information; sigmoid-transformed actor actions became muscle excitations.Source: Morphology-conditioned policy · p. 13
  • The unconditioned MLP received no morphology, while Concat encoded morphology separately and concatenated it with encoded observations at each backbone’s input.These comparisons assessed morphology conditioning and how morphology features were incorporated.Source: Morphology-conditioned policy · p. 13

Virtual model augmentation

Virtual model augmentation perturbs regional scales and total mass to expose training to model variation while retaining each original model’s reference motions.

Virtual model augmentation

  • Virtual models were generated by correlated Gaussian perturbations to regional scale factors and total mass, with shared size and limb factors creating correlated variation.Paired-region asymmetry factors changed the two sides in opposite directions, while reference motions remained fixed to the corresponding original model.Source: Virtual model augmentation · pp. 13–14
  • Standard augmentation used an asymmetry scale informed by bilateral tibial-length differences in healthy adults; asymmetry augmentation increased that scale based on clinical reports of lower-limb asymmetry.All other covariance settings were held unchanged between schemes.Source: Virtual model augmentation · p. 14
  • The study compared augmentation schemes at equal virtual-model counts per subject and separately varied the count to assess augmentation’s contribution.Training sampled subjects uniformly, then sampled uniformly among each selected subject’s original and virtual models.Source: Virtual model augmentation · p. 14

Policy training

Training optimized a morphology-conditioned actor and critic with PPO across sampled models and reference motions, using imitation rewards and termination criteria tied to tracking and episode progress.

Policy training

  • Training maximized expected discounted imitation reward combining body- and joint-pose and velocity tracking terms.Body pose targets aligned with simulated pelvis horizontal position and heading while retaining reference pelvis height; pelvis and body velocity targets used world coordinates.Source: Policy training · p. 14
  • Episodes began at sampled reference states and ended on motion completion, time-limit expiry, or a tracking-error threshold.Source: Policy training · p. 14
  • The morphology encoder, actor, and critic were jointly optimized with PPO using parallel mjlab and MuJoCo Warp simulations.Models were periodically reassigned to environment groups; the template-trained MLP used the MyoFullBody template without virtual augmentation.Source: Policy training · p. 14

Evaluation protocol

The evaluation tests policy transfer under matched held-out conditions and extends the protocol to image-derived models, unilateral limb shortening, and input perturbations. Comparisons with measured gait signals assess simulated outputs against human data.

Evaluation protocol

  • All policies were evaluated on the same 31 individual models and reference motions using fixed policy parameters and checkpoint normalization statistics.Each subject–motion pair was evaluated once with the policy mean action; trials ended at completion or when specified position or orientation error thresholds were exceeded.Source: Evaluation protocol · p. 14
  • The Gait120 experiment compared MyoAcross, Concat, and the unconditioned MLP, recording muscle activations and vertical GRFs for comparison with subject measurements.Evaluation used the common procedure and checkpoints at 15,000 completed training iterations.Source: Evaluation protocol · p. 14
  • Unilateral shortening tests changed thigh and shank lengths on each side separately while preserving model mass and the original reference motions.Standard and asymmetry-augmented policies were evaluated on the same test models, with each original model as the unmodified comparison.Source: Evaluation protocol · p. 14
  • Input sensitivity was tested by perturbing morphology features, policy reference-motion inputs, or both while keeping physical model parameters fixed.Reference perturbations changed policy targets but not initialization, reward, termination, or scoring; the MLP received only reference-motion offsets.Source: Evaluation protocol · pp. 14–15
  • The single-image experiment evaluated 88 MoVi-derived models, including 22 held-out individuals, with one standard-augmentation MyoAcross checkpoint and no individual fine-tuning.The 64 checkpoint source motions were retargeted to each model, and results were reported for all 88 models and separately for the held-out group.Source: Evaluation protocol · p. 15

Performance metrics and statistical analysis

The section defines control-performance metrics and describes how model-fit errors, simulated gait waveforms, and uncertainty intervals were computed. Aggregation gives individuals equal weight, with paired comparisons for matched conditions.

Performance metrics

  • Skeletal measurement fitting error is the mean absolute error across 17 target measurements per model, averaged over 141 models.Control metrics were averaged over motions within each individual and then equally across individuals; input-sensitivity metrics also averaged perturbation samples, while motion-category summaries used category motions.Source: Performance metrics and statistical… · p. 15
  • Completion measures the fraction of reference motions finished without early termination, while reference coverage measures the fraction of each motion executed.Tracking error is mean world-coordinate Euclidean position error across 17 rigid bodies and executed frames; it differs from the aligned body-position training-reward term.Source: Performance metrics and statistical… · p. 15

Waveform comparisons

  • EMG comparisons map 12 processed right-leg channels to named model muscles and correlate normalized gait-cycle profiles without optimizing additional phase shifts.Retained cycles lasted 0.60–1.50 s, included at least 0.08 s of right-foot non-contact samples, and were interpolated to 100 points.Source: Performance metrics and statistical… · p. 15
  • GRF profiles sum foot–ground contact forces per foot, normalize forces by body weight, and compare fully observed stance intervals interpolated to 100 points.Stance used a 0.05-body-weight threshold and minimum 0.2-s duration; intervals clipped by recording boundaries were excluded.Source: Performance metrics and statistical… · p. 15
  • Waveform profiles were averaged within motions or trials before Pearson correlations; GRF analysis also computed body-weight-normalized RMSE, with bilateral summaries requiring valid data from both feet.EMG correlations were averaged across channels within each subject, and GRF measures across feet within each subject; effective counts were reported for each signal.Source: Performance metrics and statistical… · p. 15

Statistical analysis

  • 95% percentile bootstrap intervals used 10,000 resamples of 31 subjects; completion differences used paired resamples retaining both policies’ results per subject.Intervals were formed from the 2.5th and 97.5th percentiles; paired completion intervals quantify uncertainty across subjects conditional on the trained checkpoints.Source: Performance metrics and statistical… · pp. 15–16

Related papers

Loading 2610.36535v1…