TL;DR
MyoAcross is a new framework that enables a single "generalist" neural network to control hundreds of different person-specific musculoskeletal models without any individual fine-tuning. By conditioning the control policy on body shape and mass, the system achieves high-fidelity motion tracking across diverse body types, even when the models are reconstructed from a single photograph.
What Problem Does This Paper Address?
In biomechanics and robotics, musculoskeletal models are person-specific; a tall, slender person has different limb lengths, masses, and muscle leverages than a shorter, heavier person. Traditionally, if you want to simulate how a specific individual moves using muscle-driven control, you have to train a person-specific "controller" (or policy).
This is computationally expensive and makes it difficult to perform comparative studies—for example, asking how a specific gait pattern changes across a population. The challenge is creating a "generalist" policy that can accommodate these morphological differences "zero-shot" (on models it has never seen before) while maintaining biomechanical accuracy.
Methods
The authors introduce MyoAcross, which consists of three main pillars:
- Individual Model Construction: Instead of simple uniform scaling, they use a "regional scaling" approach. By leveraging the SKEL and SMPL body models, they fit 13 regional scale factors to 17 skeletal measurements. This captures individual proportions (like the ratio of thigh to shank length) much more accurately than global scaling.
- Morphology-Conditioned Policy: The control policy (an actor-critic network) is fed "morphology features"—17 skeletal measurements plus total mass. These features are processed by a morphology encoder and injected into the network using Feature-wise Linear Modulation (FiLM). This allows the same network weights to adapt their behavior based on the specific body it is currently controlling.
- Virtual Model Augmentation: To make the policy robust, they don't just train on real subjects. They generate thousands of "virtual" models during training by applying correlated Gaussian perturbations to limb lengths and masses, exposing the network to a vast range of possible dynamics.
Datasets
- AMASS (KIT & BMLmovi): Used to build the core training (110 subjects) and evaluation (31 subjects) sets.
- Gait120: A dataset of 120 individuals used for biomechanical validation, comparing simulated results against real-world ground reaction forces (GRF) and electromyography (EMG) data.
- MoVi: Used to test the "single-image" entry point, where models were built from a single photograph of 88 different individuals.
Key Insights and Ideas
- Proportions Matter More Than Size: Regional scaling reduced skeletal fitting error by ~80-90% compared to standard global scaling. Capturing the specific "geometry" of an individual is crucial for realistic muscle-driven simulation.
- Explicit Conditioning is Key: Simply training on many models (an unconditioned MLP) isn't enough. The policy needs to know the morphology of the model it's driving. MyoAcross reduced tracking error by 22.6% compared to an unconditioned baseline.
- Asymmetry as a Stress Test: The authors tested the policy on models with unilateral lower-limb shortening (simulating conditions like cerebral palsy). By including "asymmetry augmentation" in training, the generalist policy could maintain stable gait even with a 10% difference in leg length.
- Single-Image Accessibility: One of the most exciting "aha" moments is that a model built from a single photo can be controlled by the pre-trained MyoAcross policy with 99.29% completion, potentially democratizing person-specific biomechanical simulation.
Key Results
- High Completion: MyoAcross achieved a 99.24% completion rate across 31 held-out individual models for a variety of motions (walking, turning, circling).
- Tracking Accuracy: It reduced tracking error to 53.70 mm, significantly outperforming unconditioned models.
- Biomechanical Validity: Simulated vertical ground reaction forces showed a high correlation (0.853) with real-world measurements from the Gait120 dataset.
- Muscle Activation: Simulated muscle excitations showed channel-dependent agreement with real EMG data, particularly for the soleus muscles (correlations of 0.766 and 0.825).
Limitations and Future Work
- Common Topology: The framework currently assumes all models share the same "map" of muscles and joints (topology). It cannot yet transfer between models with different numbers of muscles.
- Simplified Muscle Properties: The study assumes "geometric similarity" for muscle strength and tissue properties. In reality, individuals have varying muscle fiber types and strengths that aren't captured by bone length alone.
- Clinical Nuance: While the model handles geometric asymmetry (like limb shortening), it doesn't yet account for the neurological or sensory changes that often accompany clinical conditions.
- Future Direction: The authors suggest extending this to individually measured muscle function and integrating it with wearable robotics simulations to compare assistance strategies across different body types.
Summary updated