Isaac for Physical AI#
Isaac for Physical AI is a reference architecture for collecting teleoperated humanoid robot demonstrations to fine-tune a GR00T N policy and deploying this policy on a real or simulated robot.
The current release targets the Unitree G1 humanoid robot. A step-by-step tutorial covers a pick-and-place task (apple-to-plate), but you can use the same workflow for other tasks.
The workflow is built from the following Isaac packages:
Isaac ROS Unitree G1 Teleop Bringup — Extended Reality (XR) / interactive-marker teleoperation.
Isaac ROS Unitree G1 Recorder — MCAP rosbag recorder with session management.
Isaac ROS Unitree G1 GR00T — deploy launch for the fine-tuned GR00T policy.
Fine-tuning GR00T N and exporting the trained policy using LEAPP are performed
outside this repository using the
nvidia-isaac/gr00t-leapp-export
project (branch n17-export). Refer to the
external steps in the end-to-end tutorial below.