MuscleMimic Models

Oneline install:

pip install musclemimic-models

Load a model in two lines:

import musclemimic_models as mm

model, data = mm.load("myofullbody")   # or "bimanual"

Other helpers:

import mujoco, mujoco.viewer
import musclemimic_models as mm

# List available models
print(list(mm.REGISTRY))               # ['bimanual', 'myofullbody']

# Get the raw MJCF path
xml_path = mm.get_xml_path("bimanual")
model = mujoco.MjModel.from_xml_path(str(xml_path))

# Launch the interactive viewer
model, data = mm.load("bimanual")
mujoco.viewer.launch(model, data)

Musclemimic_models is part of the MuscleMimic research project, in which we created physiologically realistic, muscle-driven musculoskeletal models built on top of MyoSuite. This repository is designed to provide users with two musculoskeletal models: MyoBimanualArm and MyoFullBody, that could be used together or independently from the Musclemimic pipeline.

MyoFullBody enables realistic full-body motion control with pure muscle actuation. Below are example fullbody motions demonstrating the model's capabilities, all policies were trained with MuscleMimic.

MyoFullbody also allows accurate kinematics when trained with MuscleMimic on AMASS data

MyoBimanualArm focuses on upper-limb musculoskeletal control, enabling faster training convergence while preserving full finger articulation capabilities. (The videos shown below were recorded with finger actuation disabled)

Musculoskeletal Models

Both musculoskeletal models are built on MyoSuite components, combining MyoArm, MyoLegs, and MyoTorso models with Hill-type muscle actuators in MuJoCo. This enables studying motor control at the neuromuscular level and realistic muscle output, rather than via idealized joint torque controllers.

Environment Summary

$^*$ denotes configurations with finger muscles temporarily disabled.

MyoBimanualArm Environment

The MyoBimanualArm environment is designed for upper-body manipulation task. Explicit contacts are enabled in between both arms and with the thorax.

MyoFullBody Environment

The MyoFullBody environment provides a comprehensive full-body musculoskeletal system with full biomechanical detail and rich contact dynamics, suitable for locomotion, manipulation, and whole-body imitation. We explicitly enable additional collision pairs, such as leg–leg, arm–leg, foot-foot, to capture the required self-contact behavior, including bimanual interactions.

Getting Started

Prerequisites

The minimum required MuJoCo version for both models is mujoco==3.2.1. To use spec with the main Musclemimic environment, please use mujoco>=3.3.0.

Overview

The structure of the Musclemimic model is as follows. We use MyoFullBody as an example.

musclemimic_models/
└── model/
    ├── arm/
    │   ├── assets/
    │   └── myoarm_bimanual.xml
    ├── body/
    │   └── myofullbody.xml
    ├── head/
    │   └── assets/
    ├── leg/
    │   └── assets/
    ├── torso/
    │   └── assets/
    ├── meshes/
    └── scene/
└── tests/

• assets/ : includes both the kinematics chain files and the assets definition files for each body segment that its under.
• meshes/ : shared mesh files used across models for bones and skulls
• scene/ : MJCF “scene” files used in both MSK as backgrounds
• arm/, body/, head/, leg/, torso/ : model components and their associated assets/
• *.xml : MJCF model definition(s) (e.g., myofullbody.xml, myoarm_bimanual.xml)
• test/: testing files for symmetry between bodies, geoms, sites and muscle

Usage

Via Pypi

Install:

pip install musclemimic-models

Via git clone

Clone and install editable (recommended for development):

git clone https://github.com/amathislab/musclemimic_models.git
cd musclemimic_models
pip install -e .

MSK Model Refinement and Validation

Muscle Jump and Symmetry

While building MyoFullBody and MyoBimanualArm, we corrected left–right limb asymmetries and addressed several unexpected muscle-jumping behaviors. A few representative fixes are shown below.

Muscle Validation

We also cross-validate the current model using previously published cadaver studies and MRI data. A few illustrative examples are included here.

License

This project is licensed under the Apache License. See the LICENSE and NOTICE files for details.

Citation

If you use MuscleMimic in your research, please cite:

@article{Li2026MuscleMimic,
  title={Towards Embodied AI with MuscleMimic: Unlocking full-body musculoskeletal motor learning at scale},
  author={Li, Chengkun and Wang, Cheryl and Ziliotto, Bianca and Simos, Merkourios and Kovecses, Jozsef and Durandau, Guillaume and Mathis, Alexander},
  journal={arXiv preprint arXiv:2603.25544},
  year={2026}
}

Acknowledgements

The models in this repository build upon MyoSuite, an open-source musculoskeletal simulation framework.