cs.ROOct 27, 2026
SaveA generalist policy for muscle-driven control across individual musculoskeletal models
Organizations: State Key Laboratory of Tribology in Advanced Equipment, Department of Mechanical Engineering, Tsinghua University, Beijing 100084, China · Beijing Key Laboratory of Transformative High-end Manufacturing Equipment and Technology, Tsinghua University, Beijing 100084, China
Abstract
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.
Figures & tables
Figure 1: Muscle-driven control across individual musculoskeletal models. a , Four example models illustrate variation in body dimensions, body mass and bilateral symmetry: A, 1.70 m and 65 kg; B, 1.86 m and 73 kg; C, 1.62 m and 75 kg; and D, 1.67 m and 65 kg, with 10% unilateral lower-limb shortening on the right side. Colours identify the same models throughout the figure. b , Motion sequences illustrate reuse of the pretrained MyoAcross policy across the four models for slow, medium-speed and fast walking, backward walking, left and right turns, and counterclockwise and clockwise circling. Each sequence shows four frames per model. Small coloured curves indicate the corresponding movement paths. CCW, counterclockwise; CW, clockwise.
Figure 2: Overview of the MyoAcross framework. a , The 64 source motions comprise 8 clips in each of 8 locomotion categories. CCW, counterclockwise; CW, clockwise. b , SKEL derives target skeletal measurements from SMPL body shape. Skeletal-guided regional scaling and mass calibration adapt the MyoFullBody template to construct 141 individual musculoskeletal models. Retargeting the source motions to each model produces 9,024 reference motions. c , Skeletal measurement fitting error across the 141 models. Template, Global fit and Regional fit denote the unchanged MyoFullBody template, fitting a single global scale factor and skeletal-guided regional scaling, respectively. Each point represents one model’s mean absolute error across 17 target skeletal measurements. Boxes span the interquartile range, central lines indicate medians, and whiskers extend to the most extreme values within 1.5 times the interquartile range. The inset enlarges the distribution for regional scaling. d , Original models of 110 training individuals and their virtual variants support policy learning in parallel simulations; 31 individuals are held out for evaluation. A morphology encoder maps model-derived morphology features to a learned morphology representation, which conditions the actor and critic through feature-wise linear modulation (FiLM). The actor generates muscle excitations, and training uses proximal policy optimization (PPO).
Figure 3: Policy reuse across held-out individual musculoskeletal models. a , Training schemes and shared evaluation protocol for the template-trained MLP (labelled Template MLP), unconditioned MLP, Concat and MyoAcross. The unconditioned MLP is trained across individual models without morphology input. The three policies trained across models share the same training model bank, reference motions and interaction budget. Each policy is evaluated at 15,000 training iterations on 31 held-out individual models, with 64 corresponding reference motions per model and no individual fine-tuning. b , Completion, reference coverage and tracking error. Faint points represent individual-model means over the 64 reference motions; diamonds indicate means across individuals. Detail views enlarge completion and reference coverage for the three policies trained across models. c , Paired comparisons of individual-model tracking error. Each point compares MyoAcross with the unconditioned MLP or Concat for the same individual. Dashed lines indicate equal error; points below these lines indicate lower error for MyoAcross. d , Tracking error across eight motion categories for the unconditioned MLP, Concat and MyoAcross. Each category contains eight motions per model; symbols indicate means across individuals. Dotted, dashed and solid lines connect category means for the unconditioned MLP, Concat and MyoAcross, respectively, as guides to the eye. Error bars in b and d show 95% percentile bootstrap confidence intervals from 10,000 subject-level resamples of the 31 individuals. Tracking error is measured in world coordinates without alignment and includes executed frames from incomplete trials.
Figure 4: Biomechanical assessment of walking across individual musculoskeletal models. a , Workflow from labelled motion-capture markers and recorded body mass to individual musculoskeletal simulation. MoSh++ estimates SMPL-H body shape from the markers. SKEL-derived skeletal measurements guide regional scaling of the MyoFullBody template, and mass calibration matches recorded body mass. Model-derived morphology features condition the pretrained policy, which generates muscle excitations to track reference motions. Eight medium-speed KIT walking motions are retargeted to each of 120 Gait120 individual models. All policies are evaluated at 15,000 training iterations, with policy parameters and normalization statistics fixed and without individual fine-tuning. b , Subject-level tracking error for MyoAcross, Concat and the unconditioned MLP ( subjects per policy). c , Simulated vertical ground reaction force (GRF) profiles for MyoAcross and measured level-walking profiles from the corresponding Gait120 subjects, shown separately for the left and right feet. Forces are normalized by body weight (BW), and complete stance phases are interpolated to a common phase grid. Lines and shaded bands show population means and standard deviations, respectively ( subjects for the left foot and 107 for the right). d , Subject-level GRF agreement, quantified by Pearson correlation (left) and BW-normalized root-mean-square error (RMSE; right). Metrics are calculated separately for each foot and averaged within subjects with valid bilateral measurements ( per policy). e , Subject-level agreement between simulated muscle activations and measured electromyography (EMG), expressed as the mean Pearson correlation across 12 EMG channels ( per policy). In b, d and e, points represent individual subjects. Boxes span the interquartile range, central lines indicate medians, and whiskers extend to the most extreme observations within 1.5 times the interquartile range. Waveform correlations are calculated within each subject before population aggregation. No correlation-maximizing temporal shifts are applied.
Figure 5: Virtual model augmentation supports policy reuse across individual and asymmetric models. a , Training schemes and evaluation protocols. V0 uses 110 original training models without virtual model augmentation. Standard augmentation (V10) and asymmetry augmentation each add 10 virtual models per training individual, yielding 1,210 training models. Asymmetry augmentation increases the asymmetry perturbation scale while retaining the other covariance settings. All schemes use the MyoAcross policy architecture and the same interaction budget, with evaluation at 15,000 training iterations on models of 31 held-out individuals and 64 reference motions per individual. b , Completion, reference coverage and tracking error on the original held-out models for V0 and V10 (standard augmentation). Faint points represent individual-model means over the 64 reference motions; squares (V0) and circles (V10) indicate means across individuals. Out-of-range points appear as left- or right-pointing triangles at the corresponding axis limits, labelled with their measured values. c , The same metrics for standard augmentation and asymmetry augmentation under unilateral lower-limb shortening. Thigh and shank regional scale factors are reduced together by 1%, 2.5%, 5% or 10% on either the left or right side, giving nine conditions including the original model at 0%. Each test model retains its corresponding original model’s reference motions. Circles and dashed lines denote standard augmentation (labelled Standard); triangles and solid lines denote asymmetry augmentation. Symbols indicate means across individuals. Error bars in b and c show 95% percentile bootstrap confidence intervals from 10,000 subject-level resamples of the 31 individuals. Tracking error includes executed frames from incomplete trials.
Figure 6: Policy sensitivity to morphology feature and reference-motion input errors. a , Input perturbation design for the unconditioned MLP, Concat and MyoAcross, evaluated on 31 held-out individual models with 64 reference motions per model. Physical models, policy parameters and normalization statistics remain fixed. Perturbations affect only the inputs supplied to the policy; initialization, reward computation, termination and scoring use the unperturbed reference motions. Each perturbation type includes an unperturbed baseline and three nonzero levels. b , Completion, reference coverage and tracking error under relative biases of 2%, 5% and 10% in morphology features. c , The same metrics under angular offsets of , and in lower-limb reference joint targets supplied to the policy. d , Combined morphology feature biases and reference-motion input offsets, paired as 2% with , 5% with , and 10% with . The unconditioned MLP receives no morphology features, so morphology feature perturbations leave its inputs unchanged in b , and only reference-motion input perturbations apply to it in d . Each nonzero level uses three perturbation-sign samples, with signs held fixed throughout each trial. Metrics are averaged over samples and motions within each individual, then equally across individuals. Symbols indicate these means; error bars show 95% percentile bootstrap confidence intervals from 10,000 subject-level resamples of the 31 individuals. Tracking error includes executed frames from incomplete trials.
Figure 7: Single-image model construction and policy reuse. a , Human Skeleton and Mesh Recovery (HSMR) estimates SKEL shape parameters from a single photograph. The resulting skeletal measurements are calibrated to measured height and combined with recorded body mass to construct an individual musculoskeletal model. Source motions are retargeted to the constructed model, which supplies its own morphology features. Control uses the pretrained MyoAcross policy trained with standard augmentation (V10) at 15,000 training iterations, without individual fine-tuning; the snowflake denotes frozen policy weights. The individual shown in a is an author of this study and provided informed consent for publication of the image; the face is anonymized. b , Sequential medium-speed walking poses of the image-derived model of MoVi Subject 65. Red muscles and warm-coloured bones depict the policy-controlled model; translucent purple bones depict the corresponding reference motion. Each model–reference pair is translated together for horizontal presentation. c , Completion and tracking error for 88 image-derived models, each evaluated on 64 reference motions. The cohort includes 63 individuals from the original training split (blue), 22 from the held-out split (orange) and three outside the registered training and evaluation collections (grey). Each point represents one model’s completion and mean tracking error over the 64 reference motions; the inset enlarges the high-completion region. Tracking error includes executed frames from incomplete trials.
More in this topic
Policy Learning
Policy learning develops decision rules that choose actions or interventions to maximize outcomes from data, feedback, or interaction.
See all 488 papers