NVIDIA has announced the NVIDIA Isaac™ GR00T Reference Humanoid Robot, establishing the industry’s first open reference design engineered to standardize and accelerate the development of general-purpose physical intelligence. Built on the NVIDIA Jetson Thor™ onboard compute engine and integrated with the NVIDIA Isaac GR00T open development platform, the initiative provides academic institutions and global research teams with a unified hardware blueprint and open software stack to overcome the traditionally fragmented development pipeline.
Unifying the Humanoid Development Pipeline
Historically, the development of general humanoid robots has been hindered due to fragmented processes that included hardware assembly, data collection, simulation, and real-world application. However, with the launch of the NVIDIA Isaac GR00T Reference Humanoid Robot, all these challenges have been solved by integrating a sophisticated human-size robotic frame for body movement and manipulation (“body”) along with high-end edge computing and AI processes (“brain”).
Using a non-proprietary platform, it is possible to skip building an expensive infrastructure and concentrate directly on developing skills and reasoning processes.
“Humanoid robots will bring physical AI to the world’s largest industries, opening a multitrillion-dollar economic opportunity,” said Jensen Huang, founder and CEO of NVIDIA. “The NVIDIA Isaac GR00T Reference Humanoid Robot gives researchers a single, open platform to make breakthrough discoveries toward general-purpose physical intelligence.”
Also Read: SAP and Cyberwave Bridge the Gap Between Digital and Physical AI with Autonomous Robotics Deployment
State-of-the-Art Hardware Tailored for Physical AI
The integrated reference design features highly capable physical specifications optimized for complex real-world interaction and testing:
- Chassis and Mobility: Built upon the Unitree H2 Plus humanoid chassis, the platform stands nearly 6 feet tall, weighs 150 pounds, and features 31 degrees of freedom across its frame to mirror human-scale proportions and locomotion.
- Dexterous Manipulation: Dual Sharpa Wave tactile five-finger hands supply an additional 22 degrees of freedom, bringing the complete system to 75 degrees of freedom. This tactile infrastructure enables researchers to pioneer advanced, close-range manipulation and handling techniques.
- Full Body Control: The actuator provides torque up to 120 Newton meters for the arms and 360 Newton meters for the legs, allowing for a maximum payload of 7 kilograms and peak payload of 15 kilograms.
- Edge Compute: Edge computing and localized AI inference capabilities are provided by the NVIDIA Jetson AGX Thor™ T5000 module with its 2,070 FP4 teraflops Blackwell GPU, a 14-core Arm CPU, and 128GB of unified memory. The processor offers flexible power usage of 40-130 watts.
- Sensoring and Connectivity: The robot integrates multi-camera vision systems with a wide view head mounted stereo camera and wrist mounted cameras, together with inertial motion sensing technology. Connectivity onboard is provided by Wi-Fi 6, Bluetooth 5.2, Ethernet, USB, and audio peripherals. The battery capacity of 15Ah, or 0.972kWh allows the robot to operate for three hours before needing to recharge.
An Open-Source Software and Simulation Ecosystem
The modular architecture allows research teams to deploy the end-to-end framework or selectively integrate individual software blocks into existing workflows. The system leverages the full stack of the Isaac GR00T development platform, including:
- NVIDIA Isaac Teleop: Facilitates high-fidelity demonstration data capture required for imitation learning and policy training.
- Isaac GR00T Open Foundation Models: Delivers the baseline neural frameworks necessary for multi-task behaviors and humanoid reasoning.
- NVIDIA Isaac Sim and Isaac Lab: Provides highly precise virtual environments to safely train, evaluate, and iterate robot policies via simulation prior to physical deployment.
- NVIDIA Isaac ROS: Accelerates middleware pipelines to reliably transition simulated policies onto physical hardware.
The developer platform will also extend support to the Unitree G1 humanoid robot, broadening the availability of this open development framework across widely adopted research hardware.
Accelerating Global Academic and Institutional Collaboration
For the prompt development of cutting-edge innovations, some of the best research organizations from around the world will be adopting the new reference design, including the Allen Institute for AI (Ai2), ETH Zurich, Stanford Robotics Lab, and UCSD Advanced Robotics and Controls Laboratory. This common platform will make it possible for the ecosystem worldwide to collaborate, test, and distribute robotic capabilities at an unprecedented pace.
