Nvidia used its GTC Taipei keynote to announce the Isaac GR00T Reference Humanoid Robot, the company’s first piece of physical hardware aimed at humanoid researchers. The robot is a Unitree H2 Plus chassis fitted with Sharpa Wave five-finger hands, Nvidia’s Jetson AGX Thor T5000 compute module sitting where a brain stem might be, and the Isaac GR00T software stack pre-loaded so PhD students do not have to spend a semester wiring middleware before they can publish anything.
The body math, since this is what robotics people argue about at parties: nearly six feet tall, 150 pounds, 31 degrees of freedom across the body, 22 more split across the pair of five-finger hands, 75 total. Up to 120 newton-meters of torque in the arms, 360 in the legs, 7 kilogram rated payload per arm with a 15 kilogram peak. A 15Ah battery gets you about three hours before the robot needs to go nap on the charger. Jetson Thor on board means there is a Blackwell GPU and 128 GB of unified memory doing the on-robot inference, which is genuinely the kind of compute budget that was a server rack five years ago.
The launch partners are the part that signals where this product is actually aimed. Ai2, ETH Zurich, Stanford Robotics Center, and UC San Diego’s Advanced Robotics and Controls Laboratory are the named institutions. This is a research kit, not a factory deployment, and it is explicitly designed to standardize what humanoid researchers have been duct-taping together from mismatched Unitree, Boston Dynamics, and Agility hardware for the last two years. Available from Unitree in late 2026, no public price yet, which historically means “if you have to ask” but in this case probably means “the bill goes to the grant office.”
The strategic bit is that Nvidia is now operating one layer deeper into robotics than the chip-and-SDK stack it has been selling. Putting its name on the reference robot turns Isaac GR00T from a library into a platform, and the lock-in is the dataset every research lab generates against this hardware over the next three years. It is the same playbook that turned CUDA into a moat, run again in physical form.