How google-deepmind/mujoco_playground works
MuJoCo Playground provides robot simulation environments, RL training and playback, and workflows connecting simulation with physical hardware. The principal path is selecting a registered environment, training or playing a policy, and optionally deploying it; separate paths support teleoperation, system identification, and agentic manipulation. The README documents these capabilities, while source-index references and excerpts support selected internal links. Several task and hardware implementations were not sampled, so their internal call relationships are not inferred. The older RSS 2025 JAX/MJX environments remain a distinct legacy branch.
Main entry points: env_registry.py Environment registry · train.py Policy training · play.py Policy playback · mjx_env.py MJX environments
Components 25
Simulation and tasks
- Robot models
mujoco_playground/robots - Simulation tasks
mujoco_playground/tasks - Pick cube
mujoco_playground/tasks/manipulation/pick_cube/so101/task.py - Whole-body tracking
mujoco_playground/tasks/whole_body/tracking/task.py - Agentic tabletop
mujoco_playground/tasks/agentic/tabletop_v0 - Cartesian IK
mujoco_playground/tasks/agentic/tabletop_v0/ik_controller.py
RL learning
- Environment registry
mujoco_playground/_src/rl/env_registry.py - Policy training
mujoco_playground/rl/rsl_rl/train.py - Policy playback
mujoco_playground/rl/rsl_rl/play.py - Vector environment adapter
mujoco_playground/_src/rl/rsl_rl/vecenv_wrapper.py - Policy checkpoints
mujoco_playground/_src/rl/rsl_rl/checkpoint.py
Hardware workflows
- SO-101 teleoperation
mujoco_playground/tasks/teleop/so101 - System identification
mujoco_playground/tasks/sysid - Pick-cube deployment
mujoco_playground/tasks/manipulation/pick_cube/so101/deploy - Robot hardware drivers
mujoco_playground/robots/so101/real/device.py
Runtime foundation
- Environment wrappers
mujoco_playground/_src/wrapper/wrapper.py - Device-layer adapters
mujoco_playground/_src/wrapper/device_layer_wrapper.py - Asset resolution
mujoco_playground/_src/paths.py - Random seeding
mujoco_playground/_src/rl/rng.py - Compute selection
mujoco_playground/_src/xpu.py
RSS 2025 legacy
- MJX environments
rss2025/mujoco_playground/_src/mjx_env.py - Locomotion tasks
rss2025/mujoco_playground/_src/locomotion - Manipulation tasks
rss2025/mujoco_playground/_src/manipulation - Legacy training
rss2025/learning/train_jax_ppo.py
Other components
- Robot developer
Connections 26
- Robot developer → Environment registry: selects environment
README.md - Environment registry → Pick cube: registers task
mujoco_playground/_src/rl/env_registry.py - Environment registry → Whole-body tracking: registers task
mujoco_playground/_src/rl/env_registry.py - Environment registry → Simulation tasks: builds environments
mujoco_playground/_src/rl/env_registry.py - Policy training → Environment registry: loads environment
mujoco_playground/rl/rsl_rl/train.py - Policy training → Vector environment adapter: wraps environment
mujoco_playground/rl/rsl_rl/train.py - Vector environment adapter → Random seeding: seeds episodes
mujoco_playground/_src/rl/rsl_rl/vecenv_wrapper.py - Policy training → Compute selection: selects device
mujoco_playground/rl/rsl_rl/train.py - Policy playback → Environment registry: loads environment
mujoco_playground/rl/rsl_rl/play.py - Policy playback → Vector environment adapter: wraps environment
mujoco_playground/rl/rsl_rl/play.py - Policy playback → Policy checkpoints: loads policy
mujoco_playground/rl/rsl_rl/play.py - Robot developer → Policy training: trains policy
README.md - Robot developer → Policy playback: plays policy
README.md - Robot developer → SO-101 teleoperation: controls robot
README.md - SO-101 teleoperation → Robot hardware drivers: controls follower
README.md - Robot developer → System identification: fits dynamics
README.md - System identification → Robot hardware drivers: uses hardware logs
README.md - Robot developer → Pick-cube deployment: deploys policy
README.md - Pick-cube deployment → Robot hardware drivers: runs on hardware
README.md - Robot developer → Agentic tabletop: starts task
README.md - Agentic tabletop → Cartesian IK: uses IK controller
mujoco_playground/tasks/agentic/tabletop_v0/ik_controller.py - Environment wrappers → Device-layer adapters: exports adapters
mujoco_playground/wrapper.py - Robot developer → Legacy training: trains legacy tasks
README.md - Legacy training → MJX environments: runs MJX tasks
- MJX environments → Locomotion tasks: supports tasks
rss2025/mujoco_playground/_src/locomotion/g1/joystick.py - MJX environments → Manipulation tasks: supports tasks