NVIDIA sets a reference humanoid for labs, pairing Unitree's H2 Plus with Sharpa hands and a Jetson Thor brain

NVIDIA wants university labs to stop building their own humanoids. At its GTC Taipei event alongside Computex on June 1, 2026, the chipmaker announced the NVIDIA Isaac GR00T Reference Humanoid Robot, a fully specified research platform that combines a Unitree body, Sharpa hands, NVIDIA's Jetson Thor computer and its Isaac software stack. NVIDIA said the design will be available from Unitree in late 2026. It did not disclose a price.

The hardware is a Unitree H2 Plus humanoid, which NVIDIA describes as nearly six feet tall and 150 pounds, with 31 degrees of freedom in the body. Each arm ends in a Sharpa Wave tactile five-finger hand with 22 degrees of freedom, giving the complete robot 75 degrees of freedom. Arm joints deliver up to 120 newton metres of torque and leg joints up to 360 newton metres. The arms carry a rated payload of 7 kilograms and a peak payload of 15 kilograms.

Sensing comes from a head mounted stereo camera with a 140 degree horizontal and 102 degree vertical field of view, cameras on the wrists for close range manipulation, and an inertial measurement unit. Onboard computing runs on the NVIDIA Jetson AGX Thor T5000 module, which pairs a Blackwell GPU rated at 2,070 FP4 teraflops with a 14 core Arm CPU and 128 gigabytes of unified memory, in a power envelope configurable between 40 and 130 watts. A 15 amp hour battery holds 0.972 kilowatt hours, enough for about three hours of operation according to NVIDIA. The robot also has Ethernet, Wi-Fi 6, Bluetooth 5.2, USB, a microphone and speaker array, and an emergency stop on its remote.

The point of the package is the software around the body. NVIDIA said researchers will use Isaac Teleop to capture demonstration data, its open Isaac GR00T foundation models for reasoning and multitask behaviour, Isaac Sim and Isaac Lab for simulation, training and policy evaluation, and Isaac ROS middleware to deploy policies onto Jetson Thor for real time inference and control. The same workflow will also support Unitree's smaller G1 humanoid, and NVIDIA said a G1 reference workflow is expected soon on GitHub and Hugging Face.

Four research groups are named as early users: the Allen Institute for AI (Ai2), the Robotic Systems Lab at ETH Zurich, the Stanford Robotics Center and the Advanced Robotics and Controls Laboratory at UC San Diego. NVIDIA chief executive Jensen Huang said in the release that humanoids would bring physical AI to the largest industries and that the reference robot gives researchers one open platform. Skild AI cofounder and chief executive Deepak Pathak was also quoted, saying a reference design lets more researchers take part in frontier humanoid work.

The move extends a strategy NVIDIA has pursued in robotics for several years. It sells the training GPUs, the simulation software and the onboard computer, and publishes GR00T models as open weights, while leaving the robot bodies to partners. A reference design narrows the gap between those pieces. Labs that today buy a Unitree G1 or H1, fit third party hands and write their own integration code can instead buy a configuration that NVIDIA's own tools are tuned for.

Rivals approach the research market differently. Google DeepMind develops Gemini Robotics models and works with hardware partners such as Boston Dynamics, but does not sell a standard lab robot. Physical Intelligence publishes models rather than hardware. Hugging Face pushes low cost open hardware through its LeRobot project. Chinese makers including Unitree, AgiBot and Fourier already sell humanoids to universities, and Unitree's participation here gives it an NVIDIA endorsed product line.

For the global robotics market the reference design matters because research platforms shape what graduate students learn and which datasets get collected. If dozens of labs run the same body, hands and sensors, data and policies become easier to share and compare. That benefits NVIDIA's model and simulation business and Unitree's hardware sales. It also gives Sharpa, a newer hand maker, a large showcase.

There are open questions. NVIDIA has not given a price, a shipping date more precise than late 2026, or the export arrangements for a Chinese made robot body sold to U.S. and European institutions. The three hour battery life limits long data collection sessions. And labs that standardise on one vendor's stack may find it harder to move to rival models or chips later.