How pita-research/mjx_pita works
MJX_ZS trains and runs reinforcement-learning locomotion policies for the XG quadruped. The principal workflow is experiment selection and configuration, MJX environment construction, velocity-task simulation, PPO policy updates and checkpointing; play loads a policy and visualizes rollouts. Task mechanics are divided between a velocity component and reusable mechanical subcomponents, with EnvInfo carrying simulation state. The graph follows README capabilities and indexed source references; unsampled factory and training details are represented only where documented or indexed, so some relationships are omitted rather than inferred.
Main entry points: play.py Inference entrypoint · generic_env.py MJX environment · mechanical_base.py Mechanical facade · base.py Experiment registry
Components 27
Experiment and entrypoints
- Training entrypoint
src/train.py - Inference entrypoint
src/play.py - Experiment registry
src/cfg/experiments/base.py - XG configuration
src/cfg/xg_config.py
Environment and simulation
- Environment factory
src/envs/factory.py - MJX environment
src/envs/backends/generic_env.py - Environment state
src/envs/state.py - Typed info bus
src/envs/info.py - Info compatibility
src/envs/info_compat.py - Viewer and video
src/envs/backends/viewer.py
Task mechanics
- Component factory
src/components/factory.py - Task component API
src/components/base.py - Mechanical facade
src/components/mechanical_base.py - Velocity component
src/components/tasks/velocity/__init__.py - Velocity task rules
src/components/tasks/velocity/rewards.py - Actuator control
src/components/subcomponents/control.py - Kinematics
src/components/subcomponents/kinematics.py - Reset mechanics
src/components/subcomponents/reset.py - Reusable mechanics
src/components/subcomponents
Policy and learning
- Actor-critic networks
src/networks/mlp.py - LSTM network
src/networks/lstm.py - PPO trainer
src/training/ppo.py
Robot and task specs
- Robot specification
src/robots/base.py - Task specification
src/tasks/base.py
Other components
- Engineer
- Checkpoint files
- MuJoCo MJX
Connections 32
- Engineer → Training entrypoint: starts training
- Engineer → Inference entrypoint: runs inference
- Training entrypoint → Experiment registry: selects experiment
src/train.py - Inference entrypoint → Experiment registry: selects experiment
src/play.py - Experiment registry → XG configuration: builds config
src/cfg/experiments/xg_velocity.py - Experiment registry → Robot specification: specifies robot
src/cfg/experiments/base.py - Experiment registry → Task specification: specifies task
src/cfg/experiments/base.py - Training entrypoint → Environment factory: builds environment
src/train.py - Inference entrypoint → Environment factory: builds environment
src/play.py - Environment factory → MJX environment: constructs backend
- MJX environment → MuJoCo MJX: steps simulation
src/envs/backends/generic_env.py - MJX environment → Component factory: creates component
src/envs/backends/generic_env.py - Component factory → Velocity component: selects task
src/components/factory.py - Velocity component → Mechanical facade: extends mechanics
src/components/tasks/velocity/__init__.py - Mechanical facade → Task component API: implements interface
src/components/mechanical_base.py - Mechanical facade → Actuator control: delegates control
src/components/mechanical_base.py - Mechanical facade → Kinematics: uses kinematics
src/components/mechanical_base.py - Mechanical facade → Reset mechanics: delegates reset
src/components/mechanical_base.py - Mechanical facade → Reusable mechanics: uses subcomponents
src/components/mechanical_base.py - Velocity component → Velocity task rules: computes rewards
src/components/tasks/velocity/__init__.py - MJX environment → Environment state: returns state
src/envs/backends/generic_env.py - Environment state → Typed info bus: carries info
src/envs/state.py - MJX environment → Info compatibility: bridges info
src/envs/backends/generic_env.py - MJX environment → Viewer and video: delegates rendering
src/envs/backends/generic_env.py - Training entrypoint → PPO trainer: runs PPO
src/train.py - PPO trainer → Actor-critic networks: updates policy
- Inference entrypoint → Actor-critic networks: loads policy
src/play.py - PPO trainer → Environment state: collects rollouts
src/training/ppo.py - Inference entrypoint → Environment state: runs rollout
src/play.py - PPO trainer → Checkpoint files: saves parameters
- Inference entrypoint → Checkpoint files: loads parameters
src/play.py - LSTM network → Actor-critic networks: shares network family