How scutrobotlab/wheeled-legged_rl works
This repository trains and evaluates wheeled-legged robot policies in Isaac Lab, with play-time keyboard control, visualization, trajectory recording, and deployment export. The main workflow is task selection and configuration, simulated interaction, rollout collection, policy optimization, then checkpoint-based playback. Distinct task, terrain, state-machine, and algorithm branches are retained, alongside robot-world components and analysis tools. The README documents the user-facing workflow and capabilities; sampled source excerpts and the SOURCE INDEX clarify selected internal relationships, while unsampled links are omitted or marked unverified.
Main entry points: play.py Play CLI · env.py V14 Environment · on_constraint_policy_runner.py Policy Runner · env_cfg.py V14 Task Config
Components 26
User Workflows
- Training CLI
scripts/rsl_rl/train.py - Play CLI
scripts/rsl_rl/play.py - Keyboard Control
scripts/rsl_rl/keyboard_controller.py
Simulation Tasks
- V14 Task Config
source/agent_tasks/agent_tasks/direct/wheelbipe/wheelbipe_V14/env_cfg.py - V14 Environment
source/agent_tasks/agent_tasks/direct/wheelbipe/wheelbipe_V14/env.py - Base Task Config
source/agent_tasks/agent_tasks/direct/wheelbipe/wheelbipe25_v3/env_cfg.py - MDP Functions
source/agent_tasks/agent_tasks/manager/mdp/isaaclab - Terrain Catalog
source/agent_tasks/agent_tasks/manager/mdp/isaaclab/terrains.py - Terrain Commands
source/agent_tasks/agent_tasks/manager/mdp/terrain/terrain_command_manager.py - Terrain Mapping
source/agent_tasks/agent_tasks/manager/mdp/terrain/terrain_task_manager.py - State Machine Stack
source/agent_tasks/agent_tasks/direct/wheelbipe/state_machines/manager.py - Movement States
source/agent_tasks/agent_tasks/direct/wheelbipe/state_machines - State Machine Manager
source/agent_tasks/agent_tasks/manager/mdp/state_machine/manager.py
Policy Learning
- Policy Runner
source/agent_rl/agent_rl/rsl_rl/runners/on_constraint_policy_runner.py - PPO Task Config
source/agent_tasks/agent_tasks/direct/wheelbipe/agents/rsl_rl_ppo_cfg.py - NP3O Algorithm
source/agent_rl/agent_rl/rsl_rl/algorithms/np3o.py - BarlowTwins Policy
source/agent_rl/agent_rl/rsl_rl/modules/actor_critic_balowtwins.py - PPO Variants
source/agent_rl/agent_rl/rsl_rl/algorithms - Rollout Storage
source/agent_rl/agent_rl/rsl_rl/storage/rollout_storage.py - Policy Export
source/agent_rl/agent_rl/rsl_rl/utils/exporter.py
Robot World
- Robot Actuators
source/agent_world/agent_world/actuators - Heightfield Terrain
source/agent_world/agent_world/terrains/height_field.py
Analysis Export
- Trace Capture
scripts/utils/velocity_trace_html.py - HTML Export
scripts/utils/export_velocity_trace_html.py - Realtime Plotter
scripts/utils/realtime_plotter.py
Other components
- Engineer
Connections 27
- Engineer → Training CLI: starts training
- Engineer → Play CLI: starts playback
- Training CLI → V14 Task Config: selects task
- Play CLI → V14 Task Config: selects task
- Play CLI → Keyboard Control: enables control
- Training CLI → Policy Runner: trains with
- Play CLI → Policy Runner: runs policy
- V14 Task Config → V14 Environment: configures
source/agent_tasks/agent_tasks/direct/wheelbipe/wheelbipe_V14/env_cfg.py - V14 Task Config → Base Task Config: extends configuration
source/agent_tasks/agent_tasks/direct/wheelbipe/wheelbipe_V14/env_cfg.py - V14 Task Config → Terrain Catalog: selects terrain
source/agent_tasks/agent_tasks/direct/wheelbipe/wheelbipe_V14/env_cfg.py - V14 Environment → MDP Functions: uses task terms
- V14 Environment → State Machine Stack: configures state machines
- State Machine Stack → Movement States: selects machines
source/agent_tasks/agent_tasks/direct/wheelbipe/state_machines/manager.py - State Machine Manager → State Machine Stack: manages behavior
- V14 Environment → Terrain Commands: uses terrain profiles
- V14 Environment → Terrain Mapping: maps terrain IDs
- Terrain Catalog → Heightfield Terrain: defines terrain
source/agent_tasks/agent_tasks/manager/mdp/isaaclab/terrains.py - Policy Runner → NP3O Algorithm: optimizes with
- Policy Runner → PPO Variants: runs variants
source/agent_rl/agent_rl/rsl_rl/runners/on_constraint_policy_runner.py - NP3O Algorithm → BarlowTwins Policy: uses policy
source/agent_rl/agent_rl/rsl_rl/algorithms/np3o.py - NP3O Algorithm → Rollout Storage: writes transitions
source/agent_rl/agent_rl/rsl_rl/algorithms/np3o.py - Policy Runner → PPO Task Config: uses algorithm config
- Policy Runner → Policy Export: exports policy
- V14 Environment → Robot Actuators: drives actuators
- Play CLI → Trace Capture: records play data
- Trace Capture → HTML Export: exports HTML
- Play CLI → Realtime Plotter: shows live plots