RL baselines

CPU

• Stable-Baselines3 — tutorials/2.1_Train_SB3_Policy.ipynb (pip install -e ".[rl]")

• MyoReflex — tutorials/2.4_MyoReflex_Walk.ipynb

GPU

RSL-RL PPO on mjlab (see Installation for matching the torch build to your driver’s CUDA version):

pip install -e ".[mjlab]"
python scripts/train_mjlab.py myoLegWalk-v0 --env.scene.num-envs 1024

Always set --env.scene.num-envs explicitly (default is 1 — see Machine Learning Quick Start for why that stalls training) and re-pass it whenever you --agent.resume True a long walk run.

Default policies

Default mjlab (RSL-RL) checkpoints are hosted on myohub/myosuite-3-baselines on Hugging Face, and are downloaded automatically (cached by huggingface_hub) the first time they’re needed — by the tutorials, and by myosuite.utils.checkpoint_utils.find_checkpoint. They run on both the CPU and the mjlab backend:

python scripts/eval_mjlab_policy.py myoFingerPoseRandom-v0 --backend cpu

docs/baseline_checkpoints.md lists every default-run env id’s deterministic success rate; some published policies are unconverged snapshots. See Machine Learning Quick Start for how success is defined and evaluated.

Pretrained NPG / DEP-RL weight trees under myosuite/agents/ are not shipped in the pip package and are gitignored. Train your own policy, or use MuscleMimic checkpoints from Hugging Face (see myosuite/integrations/musclemimic/README.md).