G1 23-DOF Locomotion with Isaac Lab

This repository contains my first stab at training a locomotion policy for the Unitree G1 (23DOF) using Isaac Lab. I also implemented symmetry loss and removed the phase-based rewards. First stab at sim2sim deployment using Mujoco is also included.

The following readme.md was generated by Cursor - please forgive its slop vibe

g1-23dof-isaac-lab-current.mp4

🚀 Quick Start

Training

# Basic training
python scripts/rsl_rl/train.py --task=Loco --headless

Evaluation

# Play trained policy
python scripts/rsl_rl/play.py --task Loco --num_envs 32

Parameter Sweep

# Run parameter sweep with automatic analysis
python sweep_script_with_analysis.py

📁 Project Structure

├── scripts/rsl_rl/           # Training and evaluation scripts
│   ├── train.py             # Main training script
│   ├── play.py              # Policy evaluation script
│   └── cli_args.py          # Command line argument definitions
├── deployment/              # Sim2sim deployment
│   ├── deploy_sim.py        # MuJoCo deployment script
│   ├── policy.pt            # Trained policy weights
│   └── g1_description/      # Robot description files
├── source/                  # Isaac Lab environment source
│   └── g1_23dof_locomotion_isaac/
│       └── tasks/manager_based/g1_23dof_locomotion_isaac/
│           ├── g1_23dof_locomotion_isaac_env_cfg.py  # Environment configuration
│           ├── agents/rsl_rl_ppo_cfg.py              # PPO agent configuration
│           └── events.py                             # Domain randomization events
├── sweep_script_with_analysis.py  # Parameter sweep automation
├── sweep_analyzer.py              # Sweep results analysis
└── outputs/                      # Training outputs and logs

🔧 Configuration

Experiment Setup

When running sweep_script_with_analysis.py, ensure the experiment name is consistent across:

• sweep_script_with_analysis.py (line 8)
• source/g1_23dof_locomotion_isaac/g1_23dof_locomotion_isaac/tasks/manager_based/g1_23dof_locomotion_isaac/agents/rsl_rl_ppo_cfg.py

Current Experiment

• Experiment Name: g1_23dof_sweep_v16
• Resume from Run: 27 (configurable in sweep script)

Current Parameter Sweep

The active sweep investigates:

• env.rewards.feet_air_time.weight: [1.0, 3.0, 5.0]
• env.rewards.both_feet_air.weight: [-0.1, -0.5, 0.0]
• env.rewards.action_rate_l2.weight: [-0.005, -0.01]
• env.rewards.joint_deviation_arms.weight: [-0.1, -0.2]

🎮 Deployment

Sim2Sim Deployment

The deployment/ folder contains MuJoCo-based deployment scripts:

cd deployment
python deploy_sim.py

Keyboard Control

• WASD: Forward/backward/left/right movement
• Q/E: Yaw rotation
• Arrow Keys: Alternative control scheme

🚧 Current Challenges

1. Domain Randomization: Current sim2sim performance is poor - need to significantly expand randomization parameters (friction, mass, motor gains, terrain properties) to improve transfer robustness
2. Real-time Policy Control: Need to implement real-time policy switching/control in Isaac Sim to verify if poor performance is due to sim2sim gap vs human control limitations

📝 License

This project is based on the Isaac Lab framework and follows the same licensing terms.

---

Note: This is a work in progress. The current implementation focuses on symmetry-based learning approaches and requires further domain randomization work for robust sim2sim deployment.