How ccrprepo/amp_mjlab works
AMP_mjlab trains a unified G1 locomotion-and-recovery policy using mjlab environments and rsl_rl’s AMP/PPO stack. The main workflow configures the task, gathers simulated experience, updates the policy, and supports checkpoint playback and ONNX export. Separate velocity-task configuration and shared rsl_rl capabilities remain visible as supporting branches. Source excerpts are partial; links between unsampled components are included only where the README or SOURCE INDEX documents them.
Main entry points: amp_ppo.py AMP PPO · amp_env_cfg.py AMP task config · train.py Train entry · env_cfgs.py G1 AMP config
Components 21
Simulation tasks
- AMP task config
src/tasks/amp_loco/amp_env_cfg.py - G1 AMP config
src/tasks/amp_loco/config/g1/env_cfgs.py - AMP task logic
src/tasks/amp_loco/mdp - Velocity task config
src/tasks/velocity/velocity_env_cfg.py - Velocity task logic
src/tasks/velocity/mdp
RL learning
- AMP PPO
rsl_rl/algorithms/amp_ppo.py - PPO
rsl_rl/algorithms/ppo.py - Actor critic
rsl_rl/modules/actor_critic.py - Style discriminator
rsl_rl/modules/discriminator.py - Intrinsic reward
rsl_rl/modules/rnd.py - Observation normalization
rsl_rl/modules/normalizer.py - Rollout storage
rsl_rl/storage/rollout_storage.py
Training and playback
- Train entry
scripts/train.py - Playback entry
scripts/play.py - AMP runner adapter
src/tasks/amp_loco/rl/runner.py - AMP on-policy runner
rsl_rl/runners/amp_on_policy_runner.py
Motion and observation
- AMP motion loader
rsl_rl/utils/motion_loader.py - Task motion loader
src/tasks/amp_loco/ampmotion_loader.py - Observation history patch
mjlab_patch/mjlab/managers/observation_manager.py
Other components
- Operator
- ONNX policy
Connections 19
- Operator → Train entry: starts training
- Operator → Playback entry: starts playback
- Train entry → AMP task config: selects task
- G1 AMP config → AMP task config: customizes
src/tasks/amp_loco/config/g1/env_cfgs.py - AMP task config → AMP task logic: configures terms
src/tasks/amp_loco/amp_env_cfg.py - Velocity task config → Velocity task logic: configures terms
src/tasks/velocity/velocity_env_cfg.py - AMP task config → Observation history patch: uses history ordering
src/tasks/amp_loco/amp_env_cfg.py - Velocity task config → Observation history patch: uses history ordering
src/tasks/velocity/velocity_env_cfg.py - G1 AMP config → AMP runner adapter: selects runner
- AMP runner adapter → AMP on-policy runner: delegates to
src/tasks/amp_loco/rl/runner.py - AMP on-policy runner → AMP PPO: configures algorithm
rsl_rl/runners/amp_on_policy_runner.py - AMP PPO → Rollout storage: records transitions
rsl_rl/algorithms/amp_ppo.py - AMP PPO → Style discriminator: uses discriminator
- PPO → Rollout storage: records transitions
rsl_rl/algorithms/ppo.py - AMP PPO → Intrinsic reward: supports intrinsic reward
rsl_rl/algorithms/amp_ppo.py - PPO → Intrinsic reward: supports intrinsic reward
rsl_rl/algorithms/ppo.py - Task motion loader → AMP motion loader: motion integration
- Playback entry → ONNX policy: exports ONNX
- Train entry → ONNX policy: exports ONNX