MyoSuite 3 baseline checkpoints

Default-run mjlab (RSL-RL / PPO) policy checkpoints and evaluation videos for MyoSuite 3 the musculoskeletal simulation suite built on MuJoCo. Every checkpoint here reached at least 25% deterministic success (the mean-action policy over whole episodes; see checkpoints/README.md for the exact rates and caveats), except the MuscleMimic policy, which is scored by clip completion.

Contents

• checkpoints/<env_id>/model_<iter>.pt -- one RSL-RL checkpoint per env id, loadable on both the CPU and the mjlab (GPU) backend.
• evals/eval_<env_id>.mp4 -- a rendered evaluation rollout, for the env ids that have one.
• checkpoints/myoMimicFullbody-v0-walking_medium06/ -- the MuscleMimic full-body single-clip walking policy (mjlab only; see checkpoints/README.md).

Checkpoints are tied to the observation contract of the MyoSuite version they were trained with; see checkpoints/README.md for the compatibility note and the current success rates. Each checkpoint folder has a manifest.json (env contract hash, MyoSuite version and commit, success rate), and each MyoSuite release has a tag: use revision="v3.0" to pin the MyoSuite 3.0 baselines (MyoSuite 3.0 downloads this tag by default).

Usage

These are downloaded automatically by MyoSuite's tutorials and scripts/eval_mjlab_policy.py (find_checkpoint, see myosuite/utils/checkpoint_utils.py) whenever no local checkpoint is found -- no manual download needed if you have myosuite installed. To fetch one directly::

from huggingface_hub import hf_hub_download
path = hf_hub_download(
    "myohub/myosuite-3-baselines", "checkpoints/myoElbowPose1D6MRandom-v0/model_200.pt",
    revision="v3.0",
)

or with the CLI::

hf download myohub/myosuite-3-baselines checkpoints/myoElbowPose1D6MRandom-v0 --revision v3.0 --local-dir .

Citation and data sources

If you use these checkpoints, cite MyoSuite (see the citation section). The MuscleMimic policy also needs:

• MuscleMimic: Li et al. (2026), Towards Embodied AI with MuscleMimic: Unlocking full-body musculoskeletal motor learning at scale, arXiv:2603.25544.
• The imitated clip comes from the KIT whole-body human motion database (Mandery et al., ICAR 2015; Mandery et al., IEEE T-RO 32(4), 2016, doi:10.1109/TRO.2016.2572685) inside AMASS (Mahmood et al., ICCV 2019, arXiv:1904.03278), retargeted with GMR (Araújo et al., 2025, arXiv:2510.02252). The retargeted motions are for non-commercial research under the AMASS license; the dataset card asks you to cite AMASS and MuscleMimic.

All references: the References page of the MyoSuite documentation.

License

Apache 2.0, matching MyoHub/myosuite.