Glossary · Physics simulation
MyoSuite
Also called MyoLeg, MyoLeg26, myo_sim
An open-source set of musculoskeletal models and tasks for MuJoCo, built for research on learning muscle control. Its leg models are MyoLeg26 and MyoLeg.
MyoSuite is a collection of anatomical models for MuJoCo (hands, arms, legs and a full body), with muscles on their paths over the bones, made for training reinforcement learning agents to control muscles. Its models come from established biomechanics models (OpenSim’s), converted for MuJoCo’s speed. It is licensed Apache-2.0.
Its two leg sets:
- MyoLeg26: 13 muscles a leg on 26 paths (several muscles grouped into one), including the two-joint hamstrings, rectus femoris and gastrocnemius. From OpenSim’s gait2392 model.
- MyoLeg: 40 muscles a leg with wrapping surfaces round the bones. From the Rajagopal 2016 model.
In this project
In Muscles instead of motors both leg sets were mapped onto the fighter’s own legs, not the other way round, so his skeleton, contacts and reference motions stayed the same and the comparison with his servo body stayed direct. With a torque-neutral zero action, a joint-limit cost and smaller PPO updates, MyoLeg26 learnt the stance from scratch in 33.1–49.6 M steps and the full MyoLeg in 63.3–73.5 M, against the servo body’s 12.7–12.9 M.