Go2 Robot Training - Walk, Run, Jump

Train a Unitree Go2 quadruped robot using Reinforcement Learning (PPO) in the Genesis physics simulator.

Quick Start

1. Install Dependencies

# Create and activate conda environment
conda create -n genesis python=3.10 -y
conda activate genesis

# Install Genesis
pip install genesis-world

# Install RL library (IMPORTANT: must be this exact version)
pip install rsl-rl-lib==2.2.4

# Install other dependencies
pip install torch tensorboard

2. Train the Robot

cd /Users/aks/Desktop/Robotbuild

# Train walking (stable locomotion at 0.3-0.8 m/s)
python train.py --mode walk

# Train running (fast locomotion at 1.0-2.5 m/s)
python train.py --mode run

# Train jumping
python train.py --mode jump

# Train all modes sequentially
python train.py --mode all

Training Options

python train.py --mode walk \
    --num_envs 4096 \        # Number of parallel environments (default: 4096)
    --max_iterations 300 \   # Training iterations (default: 300)
    --viewer \               # Show visualization during training
    --resume                 # Resume from latest checkpoint

For machines without GPU:

python train.py --mode walk --cpu --num_envs 64

3. Monitor Training

tensorboard --logdir=logs/

Open http://localhost:6006 in your browser.

4. Evaluate Trained Models

# Evaluate a trained model with visualization
python evaluate.py --checkpoint logs/go2_walk_XXXXXX/

# Run more episodes
python evaluate.py --checkpoint logs/go2_run_XXXXXX/ --episodes 10

Training Modes

Mode	Speed Target	Description
walk	0.3-0.8 m/s	Stable, energy-efficient walking
run	1.0-2.5 m/s	Fast running with dynamic gait
jump	N/A	Jumping with stable landing

Project Structure

Robotbuild/
├── train.py          # Main training script
├── evaluate.py       # Model evaluation script
├── go2_env.py        # Environment with all reward functions
├── urdf/             # Robot model files
│   ├── go2/          # Go2 robot URDF
│   └── plane/        # Ground plane URDF
├── logs/             # Training logs and checkpoints
└── models/           # Saved models

Reward Functions

The environment includes comprehensive reward functions:

Locomotion rewards:

• tracking_lin_vel - Track commanded velocity
• tracking_ang_vel - Track commanded turning
• forward_vel - Reward forward motion (run mode)

Stability penalties:

• lin_vel_z - Penalize vertical bouncing
• ang_vel_xy - Penalize roll/pitch rotation
• orientation - Penalize tilting
• base_height - Maintain proper height

Smoothness penalties:

• action_rate - Penalize jerky movements
• dof_acc - Penalize joint acceleration
• energy - Penalize energy consumption

Jump-specific:

• jump_height - Reward height achieved
• air_time - Reward time in air
• land_stable - Reward stable landing

Troubleshooting

"No module named genesis"

pip install genesis-world

"Please install rsl-rl-lib==2.2.4"

pip uninstall rsl-rl rsl_rl  # Remove any old versions
pip install rsl-rl-lib==2.2.4

Training is very slow

• Use GPU: remove --cpu flag
• Reduce environments: --num_envs 1024

Robot falls immediately

• This is normal at the start of training
• Wait for 50-100 iterations for learning to begin
• Check TensorBoard for reward curves

Tips

1. Start with walking - It's the foundation for other behaviors
2. Use GPU - Training is 10-100x faster on GPU
3. Monitor TensorBoard - Watch for reward convergence
4. 300 iterations is usually enough for basic walking
5. Resume training if you need to continue: --resume