Historically, researchers developing general-purpose robots have been bottlenecked by fragmented workflows, forcing them to manually stitch together proprietary hardware, data collectors, and simulators before beginning actual AI training. During his Computex keynote on June 1, 2026, CEO Jensen Huang announced the NVIDIA Isaac GR00T reference design. This open platform provides a unified software stack designed to eliminate these exact friction points and accelerate frontier physical AI research.
The Hardware and Computing Backbone
The flagship chassis for this blueprint is the newly revealed Unitree H2 Plus. Standing six feet tall and weighing 150 pounds, the robot features 31 degrees of freedom in its base configuration. To achieve human-level dexterity, the system integrates Sharpa Wave tactile five-finger hands, bringing the total to 75 degrees of freedom across the body in Unitree’s new H2 Plus chassis.
The computing core driving this hardware is the Jetson AGX Thor T5000. Powered by an NVIDIA Blackwell GPU and a 14-core ARM CPU, the unit delivers 2,070 FP4 teraflops of AI performance. The humanoid supports multi-view sensing, whole-body control with a peak payload of 15 kilograms, and runs on a 0.972kWh battery yielding roughly three hours of continuous operation, hitting the primary goals of the Computex keynote announcement.
The platform bridges the physical gap by relying heavily on synthetic data generation via the underlying Isaac GR00T open foundation model architecture. Developers can simulate the equivalent of thousands of hours of human demonstration inside virtual environments in just a few hours.
Academic adoption occurred immediately following the announcement. A coalition of leading institutions, including the Stanford Robotics Center, ETH Zurich, Ai2, and UC San Diego’s Advanced Robotics and Controls Laboratory, formally committed to utilizing the reference design for their physical AI workflows.
What the Unified Standard Means for Physical AI
The introduction of a standardized reference design shifts the economics of robotics development. Previously, university labs spent the majority of their funding and time engineering custom locomotive hardware and building proprietary simulators just to test basic cognitive models.
By providing a baseline 75-DOF chassis paired with a 2,070-teraflop brain, NVIDIA removes the hardware engineering tax from the equation. Institutions can directly load open-source models trained in Omniverse onto the Jetson AGX Thor T5000. The Unitree H2 Plus gives them immediate physical validation of those models without requiring internal actuator development.