MS-Human-700: Whole-body Human Musculoskeletal Model

Project Page | Model Repository | msgym

Overview

MS-Human-700 is a whole-body human musculoskeletal model with anatomically detailed bodies, joints and 700 muscle–tendon units. This repository combines MuJoCo models and assets, Gymnasium environments, DynSyn-SAC training source and motion references for locomotion and right-hand manipulation.

DynSyn-SAC maps sensory features to muscle-group actions and expands them into individual muscle excitations through a learned synergy layer.

This release includes task environments, training code, and reference motions. Trained policy checkpoints and evaluation results are not included.

Related papers:

• MS-Human-700 (ICRA 2024)
• DynSyn (ICML 2024)

Models and tasks

Model XML files and their shared assets are in msgym/MS-Human-700/. Open an entry XML with MuJoCo's simulate viewer to inspect the model.

Model entry	Tasks	Muscle actuators
MS-Human-700.xml	Walking, stairs, perturbation recovery	700
MS-Human-700-FarmerWalk.xml	Walking with two fixed 5 kg dumbbells	700
MS-Human-700-Manipulation.xml	Vision-guided right-hand manipulation	81

Gymnasium environment	Task configs inDynSyn-SAC/configs/
msgym/LocomotionEnv-v1	tracking_walk.json, tracking_stairs.json, tracking_farmer.json, tracking_recovery.json
msgym/ManipulationEnv-v1	vision/manipulation_vision.json

Sensory observations

Environments return a dictionary with proprio, sensors and reference, plus channel-first RGB left_eye and right_eye when vision is enabled.

• Full body: proprioception, vestibular and tactile channels.
• Manipulation: proprioception, vestibular and tactile channels.
• Stereo vision: enabled in the supplied manipulation config and optional for locomotion. The manipulation environment also supports a nonvisual mode that includes object and target poses in the task reference.

For tracking, vision_enabled=false skips image rendering and the visual encoder; StateSACPolicy consumes the vector observations. Visual configs use VisionSACPolicy. Tracking renders 32x32 stereo images every 20 control steps; manipulation renders 128x128 images every five steps (eye_camera_size, vision_render_interval). Configs otherwise share task parameters and training budgets. Visual training uses a shared, frozen ImageNet ResNet-18 in evaluation mode, with torchvision's image preprocessing and 512-dimensional pooled features per eye, before the classification head.

Installation

Requires Python 3.12 or newer. From the downloaded repository directory:

uv venv --python 3.12
uv pip install -e ".[dynsyn]"

Models are bundled with the package; no separate model checkout is needed. On headless Linux systems, set MUJOCO_GL=egl for offscreen rendering.

Training

# Walking with stereo vision.
uv run --no-sync python DynSyn-SAC/SB3-Scripts/train.py -f DynSyn-SAC/configs/vision/tracking_walk.json

# Walking with vector observations only.
uv run --no-sync python DynSyn-SAC/SB3-Scripts/train.py -f DynSyn-SAC/configs/tracking_walk.json

# Vision-guided right-hand manipulation.
uv run --no-sync python DynSyn-SAC/SB3-Scripts/train.py -f DynSyn-SAC/configs/vision/manipulation_vision.json

Use the other task configs in the table above for stairs, fixed-load walking and recovery. Training outputs are written under logs/.

Evaluation

uv run --no-sync python DynSyn-SAC/SB3-Scripts/eval.py --log_path <training_run_directory> --num_episodes 5

Evaluation uses deterministic actions and the same observation processing as training. By default, it loads checkpoint/best_model.zip with its matching normalization statistics and writes evaluation.json. Add --video to record a video.

License

Code and models are released under the Apache-2.0 License.

Citation

If you use MS-Human-700 or msgym, please cite:

@inproceedings{zuo2024self,
  title={Self model for embodied intelligence: Modeling full-body human musculoskeletal system and locomotion control with hierarchical low-dimensional representation},
  author={Zuo, Chenhui and He, Kaibo and Shao, Jing and Sui, Yanan},
  booktitle={2024 IEEE International Conference on Robotics and Automation (ICRA)},
  pages={13062--13069},
  year={2024},
  organization={IEEE}
}