NVIDIA Isaac GR00T Reference Humanoid Robot Open Hardware and Software Design for Physical AI Research Streamlining Mechanical Assembly and Onboard Computing
The launch event held at GTC Taipei, NVIDIA announces the Isaac GR00T Reference Humanoid Robot which will act as an open hardware and software reference design allowing research in physical AI to be streamlined. According to the official announcement by NVIDIA, the system aims to bridge current fragmentation by providing a fully integrated set of hardware alongside a complete open software stack for robotics research, without proprietary platforms to democratize it.
This is the complete physical kit featuring integrated hardware and onboard computation which comprises of a physical chassis that builds upon the Unitree H2 humanoid robot (almost 6 feet tall, 150 lbs, 31 degrees of freedom in the body) combined with dual Sharpa Wave tactile 5 finger hands which add 22 degrees of freedom to reach a total of 75 degrees of freedom in the system and provide a fully dexterous system with highly dextrous manipulation capability. This allows researchers to work with human scale physical environments in real world conditions.
It allows for 120 Newton meters of arm torque and 360 Newton meters of leg torque via a whole body control system which corresponds to 7 kilograms of rated arm load and 15 kilograms peak arm payload. 15 Ah capacity with a 0.972 kWh battery allows for a runtime of 3 hours of use and onboard features include a on remote emergency stop to quickly disable the robot, plus standard Ethernet, Wi Fi 6, Bluetooth 5.2, and USB inputs along with a microphone and speaker system.
The robot also boasts on board computation in the form of an NVIDIA Jetson AGX Thor T5000 with a NVIDIA Blackwell GPU which provides 2070 FP4 teraflops of AI capability, a 14 core Arm CPU and 128GB of Unified Memory and configurable power output between 40 and 130 watts to accommodate demanding sensors, and features a multi view camera mounted on the head with 140 degrees of horizontal and 102 degrees of vertical field of view as well as cameras on the wrists to allow researchers to get better feedback for grasping objects, in addition to an IMU for tracking body movements.
It works on the Isaac GR00T software development platform to allow for simulation, training, and deployment and researchers can utilize Isaac Teleop to collect demonstration data, with the Isaac GR00T foundation models used for reasoning and learning tasks; the simulation and policy evaluation will happen in Isaac Sim and Isaac Lab before the trained model will be implemented on the robot with the help of Isaac ROS middleware.
The founder and CEO of NVIDIA, Jensen Huang stated, "Humanoids will be the worlds first fully capable physical AI that can do anything the humans can do, opening up a multi trillion dollar opportunity for us."
Ecosystem adoption and availability
Some of the worlds leading research groups including Stanford Robotics Center, ETH Zurich, Ai2, and UC San Diego have already adopted the reference design. In addition, they have made it work on the smaller Unitree G1 humanoid robot in order to help institutions share code and behaviors. The physical Isaac GR00T Reference Humanoid Robot is expected to ship by late 2026, while the workflows are slated to be released soon on GitHub and Hugging Face.
