NVIDIA Opens a Humanoid Robot and Locks In Its Compute Stack

NVIDIA opened its GTC Taipei keynote on June 1 with a humanoid that it does not actually make. The NVIDIA Isaac GR00T Reference Humanoid Robot is an open design built from a Unitree H2 Plus body, five-finger Sharpa hands, and an onboard NVIDIA Jetson Thor computer running NVIDIA’s Isaac GR00T software, and it will reach research labs through Unitree in late 2026. You can read the open humanoid robot reference design announcement in full on NVIDIA’s newsroom.

What NVIDIA supplies here is the compute and the open software that turn that hardware into something useful. The body and hands come from other companies. The body becomes a swappable part, and the brain stays with NVIDIA, along with the platform every lab will train on for years.

What the Reference Design Puts in One Box

Researchers building humanoids usually stitch a chassis from one vendor to hands from another, sensors from a third, and compute from a fourth, then spend months making the pieces talk to each other. NVIDIA’s pitch is that the stitching is done: one order, one integrated machine, one software environment.

The Body: Chassis, Hands, and Sensors

The chassis is a Unitree H2 Plus, the human-scale humanoid from China’s Unitree Robotics. It stands nearly 6 feet tall, weighs 150 pounds, and carries 31 degrees of freedom (DOF) across its joints. Dual Sharpa Wave tactile hands, each with five fingers, push the total to 75 degrees of freedom and give the machine the fine control needed to grip and manipulate objects.

The figures behind that reach are blunt. Arm torque runs up to 120 Newton-meters and leg torque up to 360, with a rated arm payload of 7 kilograms and a peak of 15. A head-mounted stereo camera sees across a 140-degree horizontal field, wrist cameras handle close work, and an inertial measurement unit (IMU) tracks motion. A 0.972-kilowatt-hour battery keeps it running for about three hours. Unitree publishes the wider Unitree H2 humanoid specifications on its own store.

The Brain: Jetson Thor and Isaac GR00T

The compute is where NVIDIA’s interest sits. Onboard is a Jetson AGX Thor T5000 module, built around a Blackwell GPU rated at 2,070 FP4 (4-bit floating point) teraflops, with a 14-core Arm processor, 128GB of unified memory, and a power draw that scales from 40 to 130 watts. That is enough to run perception, reasoning, and motor control on the robot itself rather than in a data center.

Sitting on top is the Isaac GR00T platform, NVIDIA’s open stack for capturing demonstration data, training in simulation, and deploying learned behaviors onto hardware. The split is deliberate. The body and hands come from outside vendors, and the parts that decide what the robot does come from NVIDIA.

Why NVIDIA Doesn’t Need to Build the Robot

NVIDIA’s robotics business has never been about robots. It sells the chips and the software that other companies build robots around. The Jetson AGX Thor developer kit lists at $3,499 and the T5000 module at $2,999 in volume, prices aimed at putting NVIDIA silicon inside as many machines as possible. The company details the Jetson Thor robotics compute platform on its product pages.

The reference design extends that logic to research. By packaging its compute and open models with a ready body, NVIDIA hands a lab a working humanoid on day one and, in the process, makes Isaac GR00T the environment that lab’s students train on for years. Habits formed in a university robotics center tend to follow graduates into industry.

The scale behind the push is hard to ignore. NVIDIA’s data center business has been posting record quarters on AI demand, as its record data center quarter showed, and physical AI, the term for machine intelligence that acts in the real world, is the next market it wants to anchor.

Humanoid robots will bring physical AI to the world’s largest industries, opening a multitrillion-dollar economic opportunity.

That line came from Jensen Huang, founder and chief executive of NVIDIA, in the announcement. The reference robot, he added, gives researchers a single open platform to work toward general-purpose physical intelligence.

Unitree’s Spotlight on Its IPO Day

The timing was not subtle. On the same June 1, Unitree Robotics went before the Shanghai Stock Exchange’s listing committee for an initial public offering (IPO) on the STAR Market, China’s tech-heavy board. NVIDIA’s keynote handed the company a global endorsement on the day its valuation was up for review.

• $6.2 billion target valuation for Unitree on the STAR Market, or roughly 42 billion yuan.
• 51.5% of Unitree’s revenue now comes from humanoid robots, its largest single source.
• 4.2 billion yuan the company plans to raise in the listing.
• $29,900 starting price for a standard H2, the cheapest full-size humanoid on the market.

For Unitree, being the body in NVIDIA’s reference design is a marketing coup. For NVIDIA, the choice says something blunter about hardware. A capable humanoid chassis is now cheap enough and good enough to treat as a commodity, sourced from whoever builds it best.

That is also why the platform already lists support for the Unitree G1, a smaller and even more widely used model. Swap the G1 in for the H2 Plus, and the software underneath stays the same.

Open Stack, Closed Rivals

Most of the best-known humanoid efforts are walled gardens. Tesla’s Optimus and Figure’s robots are built top to bottom by one company, with the hardware, the models, and the data held inside. Researchers outside those firms cannot easily test ideas on the machines or inspect how they work. NVIDIA took the opposite route, building on the open approach it set out with its open GR00T foundation models.

The Isaac GR00T platform is modular, so a lab can take the whole thing or bolt single pieces onto pipelines it already runs, and the researchers keep their own robot data, telemetry, and logs. The stack breaks into clear parts:

• Isaac Teleop captures high-quality human demonstration data for training.
• Isaac GR00T foundation models support reasoning, learning, and multitask behavior.
• Isaac Sim and Isaac Lab simulate, train, and test policies before real-world use.
• Isaac ROS middleware, the accelerated build of the Robot Operating System (ROS), moves trained policies onto machines.
• Onboard compute runs real-time inference and control on the robot itself.

Who Gets It, and When

NVIDIA lined up brand-name labs to launch with. Ai2, the Seattle research institute, ETH Zurich’s Robotic Systems Lab, the Stanford Robotics Center, and the Advanced Robotics and Controls Laboratory at UC San Diego all plan to use the reference design for frontier work. NVIDIA Research will use it too, to develop its own open models and frameworks.

The labs frame the appeal around access rather than raw power.

Robotics moves fastest when researchers can build on open platforms, share code and test ideas on real machines.

That came from Steve Cousins, executive director of the Stanford Robotics Center. He said the design gives students and collaborators an open humanoid with dexterous hands and onboard compute for creating, comparing, and sharing behaviors on physical hardware. The robot was one of several launches at NVIDIA’s Taipei keynote, which also pushed its consumer plans through the N1X Arm laptop chip.

On timing, the full reference robot will ship from Unitree in late 2026. The reference workflow for the smaller Unitree G1 is due sooner, on GitHub and Hugging Face, so developers can start before the hardware lands.

If the open design takes hold, the chassis under tomorrow’s research humanoids becomes as swappable as the graphics cards inside today’s servers, and NVIDIA owns the layer that never has to change. If labs decide they would rather not anchor years of work to one vendor’s compute, the late shipping window is when that hesitation will start to show.

Frequently Asked Questions

What Is the NVIDIA Isaac GR00T Reference Humanoid Robot?

It is an open humanoid robot reference design that combines a Unitree H2 Plus chassis, dual five-finger Sharpa Wave hands, an NVIDIA Jetson Thor onboard computer, and NVIDIA’s Isaac GR00T software. It is meant to give research labs one integrated machine instead of a hardware-and-software stitching job.

When Will It Be Available, and From Whom?

The full reference robot will be available from Unitree in late 2026. A separate reference workflow for the smaller Unitree G1 is expected sooner on GitHub and Hugging Face, letting developers begin in software ahead of receiving hardware.

Does NVIDIA Build the Robot’s Body?

No. The body is a Unitree H2 Plus chassis and the hands come from Sharpa. NVIDIA supplies the Jetson Thor compute module and the Isaac GR00T software platform, which is the part of the system it controls and intends to standardize.

How Powerful Is the Onboard Computer?

The robot runs a Jetson AGX Thor T5000 module with a Blackwell GPU rated at 2,070 FP4 teraflops, a 14-core Arm CPU, and 128GB of unified memory, drawing between 40 and 130 watts. That allows perception, reasoning, and motor control to run on the robot rather than in the cloud.

Which Research Institutions Will Use It?

Launch partners include Ai2, ETH Zurich, the Stanford Robotics Center, and UC San Diego’s Advanced Robotics and Controls Laboratory. NVIDIA Research will also use the design to advance its own open models, frameworks, and hardware.