How clearlab-sustech/wheel-legged-gym works
Wheel-Legged Gym provides Isaac Gym locomotion environments and an on-policy reinforcement-learning workflow. Users select a registered task, train or play a policy, and obtain checkpoints or policy exports. The map covers base and VMC environment variants, terrain generation, PPO policies and rollout storage. The README documents the user workflow and task capabilities; some internal connections are supported only by the source index, so uncertain links are omitted.
Main entry points: legged_robot.py Legged robot · on_policy_runner.py On-policy runner · __init__.py Task registration · task_registry.py Task registry
Components 27
Task entry
- Train command
wheel_legged_gym/scripts/train.py - Play command
wheel_legged_gym/scripts/play.py - Task registry
wheel_legged_gym/utils/task_registry.py - Task registration
wheel_legged_gym/envs/__init__.py
Simulation environments
- Base config
wheel_legged_gym/envs/base/legged_robot_config.py - Wheel config
wheel_legged_gym/envs/wheel_legged/wheel_legged_config.py - VMC config
wheel_legged_gym/envs/wheel_legged_vmc/wheel_legged_vmc_config.py - Flat config
wheel_legged_gym/envs/wheel_legged_vmc_flat/wheel_legged_vmc_flat_config.py - Base task
wheel_legged_gym/envs/base/base_task.py - Legged robot
wheel_legged_gym/envs/base/legged_robot.py - VMC robot
wheel_legged_gym/envs/wheel_legged_vmc/wheel_legged_vmc.py - Terrain generator
wheel_legged_gym/utils/terrain.py - Wheel-leg asset
resources/robots/wl/urdf/wl.urdf
Policy learning
- On-policy runner
wheel_legged_gym/rsl_rl/runners/on_policy_runner.py - Vector environment
wheel_legged_gym/rsl_rl/env/vec_env.py - PPO optimizer
wheel_legged_gym/rsl_rl/algorithms/ppo.py - Actor critic
wheel_legged_gym/rsl_rl/modules/actor_critic.py - Sequence policy
wheel_legged_gym/rsl_rl/modules/actor_critic_sequence.py - Recurrent policy
wheel_legged_gym/rsl_rl/modules/actor_critic_recurrent.py - Rollout storage
wheel_legged_gym/rsl_rl/storage/rollout_storage.py
Runtime support
- Runtime helpers
wheel_legged_gym/utils/helpers.py - Robot math
wheel_legged_gym/utils/math.py - Training logger
wheel_legged_gym/utils/logger.py
Other components
- Engineer
- Isaac Gym
- Model checkpoint
- TensorBoard
Connections 31
- Engineer → Train command: starts training
README.md - Engineer → Play command: runs policy
README.md - Train command → Task registry: selects task
wheel_legged_gym/scripts/train.py - Play command → Task registry: selects task
wheel_legged_gym/scripts/play.py - Task registration → Task registry: registers tasks
wheel_legged_gym/envs/__init__.py - Task registration → Legged robot: registers class
wheel_legged_gym/envs/__init__.py - Task registration → VMC robot: registers class
wheel_legged_gym/envs/__init__.py - Task registration → Wheel config: registers config
wheel_legged_gym/envs/__init__.py - Task registration → VMC config: registers config
wheel_legged_gym/envs/__init__.py - Task registration → Flat config: registers config
wheel_legged_gym/envs/__init__.py - Wheel config → Base config: inherits config
wheel_legged_gym/envs/wheel_legged/wheel_legged_config.py - VMC config → Wheel config: inherits config
wheel_legged_gym/envs/wheel_legged_vmc/wheel_legged_vmc_config.py - Flat config → VMC config: inherits config
wheel_legged_gym/envs/wheel_legged_vmc_flat/wheel_legged_vmc_flat_config.py - Task registry → Legged robot: creates environment
wheel_legged_gym/utils/task_registry.py - Task registry → On-policy runner: creates runner
wheel_legged_gym/utils/task_registry.py - Legged robot → Base task: extends task base
wheel_legged_gym/envs/base/legged_robot.py - VMC robot → Legged robot: extends robot
wheel_legged_gym/envs/wheel_legged_vmc/wheel_legged_vmc.py - Legged robot → Terrain generator: uses terrain
wheel_legged_gym/envs/base/legged_robot.py - Legged robot → Robot math: uses math
wheel_legged_gym/envs/base/legged_robot.py - Legged robot → Isaac Gym: steps simulation
wheel_legged_gym/envs/base/legged_robot.py - On-policy runner → Vector environment: steps environment
wheel_legged_gym/rsl_rl/runners/on_policy_runner.py - On-policy runner → PPO optimizer: runs algorithm
wheel_legged_gym/rsl_rl/runners/on_policy_runner.py - On-policy runner → Actor critic: selects policy
wheel_legged_gym/rsl_rl/runners/on_policy_runner.py - On-policy runner → Sequence policy: selects policy
wheel_legged_gym/rsl_rl/runners/on_policy_runner.py - PPO optimizer → Rollout storage: uses rollouts
wheel_legged_gym/rsl_rl/algorithms/ppo.py - PPO optimizer → Actor critic: uses policy
wheel_legged_gym/rsl_rl/algorithms/ppo.py - Recurrent policy → Actor critic: extends policy
wheel_legged_gym/rsl_rl/modules/actor_critic_recurrent.py - On-policy runner → Model checkpoint: saves model
README.md - Play command → Model checkpoint: loads model
README.md - On-policy runner → TensorBoard: logs metrics
README.md - Legged robot → Wheel-leg asset: loads robot asset