Closing the Sim-to-Real Gap: Training Spot Quadruped Locomotion with NVIDIA Isaac Lab

Training Spot Quadruped Locomotion in Isaac Lab

• This section explains how to train a locomotion RL policy for Spot in Isaac Lab, utilizing domain randomization to improve generalization.
• It covers defining observation and action spaces, training the RL policy, and exporting the trained model.

Play the Trained Policy

• Once the RL policy is trained, it can be played in Isaac Sim and exported to .onnx format for deployment on hardware.

Deploying the Trained RL Policy on Spot with Jetson Orin

• The deployment of trained models from simulation to the real world is challenging but made easier with the Spot RL Researcher Kit.
• The accurate physics simulation in Isaac Lab allows for zero-shot deployment on the real Spot robot with Jetson Orin for similar performance.

Deployment Steps

1. Transfer trained model and configuration files to the Jetson Orin device.
2. Move necessary files to the appropriate directories on the Jetson Orin for running the policy.
3. Run the policy on the real Spot robot using the Spot Python SDK and gamepad controls.

Conclusion

• By leveraging the capabilities of Isaac Lab and the Spot RL Researcher Kit, users can bridge the sim-to-real gap for quadruped locomotion training and deployment.
• The trained policies can be seamlessly transferred from simulation to the real world with minimal adjustments, showcasing the power of this integrated solution.