This repository provides a complete workflow for training and deploying control policies for the Unitree Humanoid G1 using two independent (but complementary) learning pipelines:
- LeggedGym → Reinforcement Learning (RL) using NVIDIA IsaacGym
- BeyondMimic → Whole-body imitation learning using IsaacLab
You can choose to train with:
- LeggedGym only (RL locomotion)
- BeyondMimic only (whole-body tracking)
- Both pipelines for a more complete humanoid control system
LeggedGym provides:
- GPU-accelerated reinforcement learning
- Massive parallel simulation (thousands of environments)
- Domain randomization for robustness
- Export of learned policies to ONNX or TorchScript
- Fast prototyping for locomotion controllers
BeyondMimic enables:
- Whole-body human-to-humanoid motion retargeting
- Reference motion tracking and imitation learning
- High-fidelity physics using MuJoCo
- Connection with IsaacLab controllers and policy runners
This repository documents:
- SIM2SIM transfer between IsaacGym → MuJoCo → IsaacLab
- Real deployment on Unitree G1 using the SDK or ROS 2
- Motor PD tuning, sensor filtering, safety layers
- Fallback control strategies for real robot experiments
A complete documentation website is included in the docs/ folder and can be published as GitHub Pages.
RL/ ├── LeggedGym/ # Training environments for legged robots (IsaacGym) ├── BeyondMimic/ # Whole-body tracking & imitation learning (IsaacLab) ├── docs/ # GitHub Pages documentation website │ ├── index.md # Website homepage (Overview + navigation) │ ├── leggedgym.md # Full LeggedGym documentation │ ├── beyondmimic.md # Full BeyondMimic documentation │ ├── sim2real.md # SIM2SIM + SIM2REAL pipeline │ ├── installation.md # Installation + environment setup (IsaacGym/IsaacLab) │ └── _config.yml # GitHub Pages configuration (theme + navbar) └── README.md # Main repository README (project overview)