BAAI open-sourced MS-Human-700, a 700-muscle human musculoskeletal model for reinforcement learning
BAAI released MS-Human-700, a whole-body human musculoskeletal model with 700 muscle–tendon units, on Hugging Face under Apache-2.0. The repository includes MuJoCo models, Gymnasium environments, DynSyn-SAC training code, and reference motions for locomotion and right-hand manipulation. Trained policy checkpoints and evaluation results are not included.
China context
- Original name
- 北京智源人工智能研究院
- Outside China
- Open weights · huggingface.co
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers outside China can access the model and code on Hugging Face under Apache-2.0, enabling integration into robotics and simulation projects without licensing restrictions.
- For investors
- The open-source release may reduce the moat for proprietary musculoskeletal simulation tools, potentially affecting companies that sell such software or services.
BAAI released MS-Human-700, a whole-body human musculoskeletal model with anatomically detailed bodies, joints, and 700 muscle–tendon units. The Hugging Face 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. The release includes task environments, training code, and reference motions, but trained policy checkpoints and evaluation results are not included. The code and models are released under the Apache-2.0 license.
MS-Human-700 provides a detailed musculoskeletal model with 700 muscle–tendon units, supporting tasks such as walking, stairs, perturbation recovery, and vision-guided right-hand manipulation. The DynSyn-SAC algorithm uses a learned synergy layer to map sensory features to muscle-group actions, then expands them into individual muscle excitations. Environments return proprioceptive, vestibular, tactile, and optional stereo vision observations. Visual training uses a shared, frozen ImageNet ResNet-18 in evaluation mode with 512-dimensional pooled features per eye. The model requires Python 3.12 or newer and supports headless rendering via MUJOCO_GL=egl.
Researchers and developers working on biomechanical simulation and reinforcement learning for humanoid control gain access to a detailed open-source musculoskeletal model and training code, reducing the cost and effort of building such environments from scratch. The Apache-2.0 license permits commercial use, which may accelerate adoption in robotics and rehabilitation research.
The open-source release under Apache-2.0 allows companies and research institutions to use and modify the model and code without licensing fees, potentially lowering barriers to entry for humanoid robotics and biomechanics projects.
The release does not include trained policy checkpoints or evaluation results, so independent verification of performance is not yet possible. A specific signal to watch is whether BAAI or the community publishes benchmark results or trained policies for the included tasks.