NVIDIA unveils an open humanoid reference design built on Unitree and Jetson Thor.

NVIDIA introduced the Isaac GR00T Reference Humanoid Robot, its first open humanoid reference design, built on the Jetson Thor compute module and the Isaac GR00T software platform. The design integrates a Unitree H2 humanoid body, dual Sharpa Wave five-finger tactile hands and NVIDIA's onboard AI stack, for a combined 75 degrees of freedom. The reference robot is aimed squarely at universities and research labs, with early adopters including Ai2, ETH Zurich, the Stanford Robotics Center and UC San Diego.

The hardware headline is Jetson Thor: an AGX Thor T5000 built on NVIDIA's Blackwell GPU delivers roughly 2,070 FP4 teraflops of AI performance with 128GB of unified memory in a 40-to-130-watt envelope — enough on-robot compute to run modern vision-language-action models in real time. That is the class of edge silicon the category has been waiting for, and packaging it into a documented, reproducible platform lowers the barrier for labs that previously had to bootstrap their own hardware.

Strategically, the move is classic NVIDIA: rather than build robots, it seeds the ecosystem's toolchain and makes its compute the default substrate. Choosing Unitree as the reference body — announced around the same window as Unitree's IPO clearance — also cements the Chinese firm's hardware as the research world's low-cost standard. For commercial buyers, a common reference design means faster maturation of the underlying models and a larger talent pool trained on NVIDIA's stack, which should compound into better fielded systems over the next few product cycles. The reference design is slated to ship in October 2026.