MuJoCo Backend

MuJoCo is the default backend path in the committed owner configs. The Python dependency is mujoco-uni==3.8.0 in pyproject.toml, and the adapter lives under src/unilab/base/backend/mujoco/.

When To Use It

  • You want the default training route for PPO, APPO, off-policy SAC/TD3, or FlashSAC.

  • The task owner exists only as conf/.../<task>/mujoco.yaml.

  • You need MuJoCo-specific tooling such as scripts/play_viser.py or scene export from a MuJoCo XML/MJB model.

Commands

uv run train --algo ppo --task go2_joystick_flat --sim mujoco
uv run train --algo appo --task go1_joystick_flat --sim mujoco training.no_play=true
uv run train --algo sac --task g1_walk_flat --sim mujoco

Playback mode is resolved by the backend contract in src/unilab/base/backend/base.py. MuJoCo reports physics-state playback support in src/unilab/base/backend/mujoco/backend.py; auto playback records video rather than opening the Motrix native interactive renderer.