Nvidia's latest humanoid move is not a finished consumer robot. At GTC Taipei, Jensen Huang showed the NVIDIA Isaac GR00T Reference Humanoid Robot, a research blueprint meant to standardize the body, hands, compute and software stack.
The reference system uses Unitree's H2 Plus body, Sharpa five-finger tactile hands, 75 degrees of freedom, stereo and wrist cameras, IMUs, a battery rated around three hours and Jetson AGX Thor T5000 with Blackwell GPU compute.
The stack matters more than the shell
The hardware is only one layer. Nvidia is packaging Isaac Teleop for data collection, Isaac GR00T for the foundation model, Isaac Sim and Isaac Lab for simulation and training, and Isaac ROS for deployment. The goal is to make labs compare methods on a common platform instead of rebuilding every link.
Stanford's Steve Cousins described it as an open humanoid reference design with dexterous hands and onboard AI compute. ETH Zurich, AI2 and UC San Diego are among the early research users named around the program.
Nvidia sells the road, not the robot
The strategy resembles Nvidia's role in model training: it does not need to build every robot if the chips, simulation tools and training pipeline depend on its platform. Huang framed humanoids as a route to physical AI in multitrillion-dollar industries.
The limits are clear. The Unitree-based robot is expected to ship toward the end of 2026, the GR00T workflow for Unitree G1 has not landed on GitHub and Hugging Face yet, and the target is academic research, not home chores. Nvidia is betting that whichever robot wins, the training base underneath can still be Nvidia's.
Sources: NVIDIA Newsroom, PR Newswire, CNBC, Engadget, CocoLoop; checked Unitree H2 Plus, Sharpa hands, 75 degrees of freedom, Jetson Thor T5000 and the stated academic research positioning.