Gym environments where the agent is a pile of muscles

What it does MyoSuite is a collection of biomechanical simulation environments built on MuJoCo and wrapped in the OpenAI Gym API. The tasks range from simple elbow poses to full-body movements, all controlled by detailed muscle-actuated models rather than the direct torque joints typical in robotics sims. It ships with pre-trained baselines, Colab tutorials from ICRA 2023 and 2024, and a “MyoSkeleton” model that can be initialized separately.

The interesting bit The twist is the fidelity of the musculoskeletal modeling: tendons, muscle wrapping, and contact-rich dynamics mean your RL agent must solve problems biology actually faces, like co-contraction and redundancy. The README calls it “contact-rich” for a reason — this is not the clean torque-controlled pendulum your Q-table learned on.

Key highlights

• Native Gym API: drop into existing RL pipelines with gym.make('myoElbowPose1D6MRandom-v0')
• MuJoCo-native rendering on macOS via mjpython and launch_passive
• Pre-trained baselines and ICRA tutorial notebooks available
• Apache 2.0 licensed, actively maintained with Slack and Twitter presence
• Python 3.10+ required; uv is the documented fast path for install

Caveats

• The “MyoSkeleton” model requires a separate initialization step (myoapi_init or python -m myosuite_init) whose details are only hinted at in the README
• macOS users must remember the mjpython wrapper for visualization, a friction point the authors explicitly flag

Verdict Grab this if you’re doing motor-control research, prosthetics, or just want to punish your RL agent with biologically plausible physics. Skip it if you need generic robotics sims — MuJoCo’s standard humanoid is less fuss and fewer tendons.