How hellod035/leggedlab works
Project type and purpose
• LeggedLab is a robotics simulation & reinforcement-learning framework built as an IsaacLab extension.
• It provides:
• Environment definitions for legged robots (Unitree G1, H1; FFT-AI gr2)
• Asset management (USD files, configurations)
• MDP logic (rewards, observations)
• CLI scripts for train/play workflows
• Integration with IsaacSim/IsaacLab and RSL_RL for multi-GPU/node RL.
File structure
Top-level modules:
• legged_lab/assets → USD robot assets & configs
• legged_lab/envs → Env definitions & configs per robot
• legged_lab/mdp → Reward functions
• legged_lab/terrains → RayCaster & terrain generator configs
• legged_lab/utils → CLI arg parsing, scene setup, keyboard, task registry
• legged_lab/scripts → train.py & play.py entrypoints
External config at root: setup.py, flake8, pre-commit, LICENSE, README
README for insights
• This is an extension to IsaacLab, requiring IsaacSim 4.5 / IsaacLab 2.1.
• Uses RSL_RL for RL back‐end; supports TensorBoard logging, multi-GPU/node.
• Workflows:
• train.py to launch RL training (gym-style envs + IsaacSim headless)
Main components & their relationships
a. User Interface
• CLI scripts (train.py / play.py)
• Accept --task, --num_envs, --logger flags
b. LeggedLab Extension (Python package)
• Task registry loads the selected env
• Env factories in legged_lab/envs create Gym-style envs
• MDP (legged_lab/mdp) defines reward & observation logic
• Terrain module (legged_lab/terrains) generates/loads terrain
• Asset module (legged_lab/assets) loads USD robot models & configs
• Utils set up the IsaacSim scene, configure physics, sensors
c. IsaacSim + IsaacLab Core
• Simulation engine that runs physics, rendering, sensor pipelines
• Exposes Python API (omni.isaac) to load USD, step sim
d. Reinforcement‐Learning Back-end
• RSL_RL orchestrates policy optimization (PPO/DPO)
• Handles multi-GPU / multi-node distribution
• Exchanges action/observation tensors with the env
e. Logging & Visualization
• TensorBoard, optionally console output
f. (Optionally) Real-robot deploy code (separate repo)
• Uses same robot USD → real robot mapping
Architectural patterns & design principles
• Modular, plugin‐style IsaacLab extension
• Factory pattern for env creation (legged_lab/envs)
• Config-driven asset & env setup (YAML + USD)
• Gym API compatibility → decouples env from RL back-end
• Separation of concerns: assets | env logic | MDP | RL training | CLI