Santa Clara, California.
For years, robotics PhD students at Stanford or ETH Zurich would spend the first 18 months of a 4-year grant building robots rather than training them. They had to source actuators from one vendor, simulation software from another, and inference hardware from a third. The result was often what insiders call a “Franken-robot”: good enough to publish a paper, but not useful for the next researcher who inherited the codebase. The NVIDIA Isaac GR00T Reference Humanoid Robot, unveiled at GTC Taipei on June 1, 2026, is NVIDIA’s direct response to this problem.
What the NVIDIA Isaac GR00T Reference Humanoid Robot Actually Is
This platform is not a typical consumer product line. Instead, it is a validated, open blueprint—a reference design that provides research institutions with a complete, working system rather than just a list of parts. NVIDIA, Unitree, and Singapore-based Sharp introduced the Isaac GR00T Reference Humanoid Robot at GTC Taipei as the first open humanoid reference design. It pairs a Unitree H2 Plus body with Jetson Thor computing and the Isaac GR00T software stack.
Top research institutions such as AI2, ETH Zurich, Stanford Robotics Center, and UC San Diego’s Advanced Robotics and Controls Laboratory plan to use this reference design to advance humanoid robotics research. This kind of institutional support is important. When places like Stanford and ETH Zurich agree on the same hardware, the field finally gets something new: reproducible experiments on a standard platform.
Inside the Chassis: The Unitree H2 Plus Specs That Define the Platform
The hardware is built around a Unitree H2 Plus chassis that stands almost six feet tall and weighs 150 pounds, with 31 degrees of freedom throughout the body. Attached to it are two Sharpa Wave tactile five-finger hands, each supplying 22 degrees of freedom, for a total of 75 across the whole system. Each fingertip has tactile sensors, which enable the precise manipulation required for activities such as using tools or assembling components.
These actuator numbers are important. The legs can produce 360 Nm of torque, which enables a humanoid robot to recover from a slip on a warehouse floor. For comparison, a person pushing hard against something stationary generates about 250 Nm of peak torque through the hip. The H2 Plus goes 44 percent beyond that—not because NVIDIA expects warehouse use right away, but because research needs to test robots at the limits of what they can do.
The robot’s sensors include a head-mounted stereo camera with a 140-degree horizontal and 102-degree vertical field of view, wrist cameras for close-up work, and an inertial measurement unit for tracking movement. This setup lets the robot track its own hands against the background simultaneously, which is necessary for assembly tasks where a person would naturally look at their fingers.
The Brain: Jetson AGX Thor T5000 and the Case for Local Inference
The computing side is where physical AI goals meet applied engineering. The robot uses an NVIDIA Jetson AGX Thor T5000 module, which has a Blackwell architecture GPU delivering 2,070 FP4 teraflops, a 14-core Arm CPU, and 128 GB of unified memory. This provides enough power to run language-based manipulation commands locally without sending data to the cloud, at the fast speeds real-time robot control requires.
This detail is important. Robots that rely on the cloud have delays that do not work with fast, quick movement. For example, a humanoid stepping over uneven ground cannot wait 80 milliseconds for a data center to send back a correction. The Jetson AGX Thor T5000 has a flexible power range from 40 to 130 watts for immediate sensor processing and robot inference. This pliability lets labs use less power during slow tabletop experiments and increase it for movement trials, which helps extend battery life without changing hardware.
The robot connects through Ethernet, Wi-Fi 6, Bluetooth 5.2, and USB, and it has microphones and speakers for voice interaction. Its battery has a 15 Ah (0.972 kWh) capacity, giving about 3 hours of use. There is also a remote emergency stop to quickly and safely turn off the robot if needed.
The Software Stack That Makes Open Physical AI Viable
The hardware specs are not the only reason five major institutions signed on before the platform even shipped. The Isaac GR00T platform also includes NVIDIA Isaac Teleop for recording high-quality robot demonstration data; Isaac GR00T open base models for humanoid reasoning and multi-task behavior; Isaac Sim and Isaac Lab for emulating and testing robot policies before real-world use; and fast Isaac ROS middleware to transfer trained policies to physical robots.
This setup solves a problem that has long divided robotics research. For example, a team at UC San Diego can record a physical demonstration, simulate it at scale in Isaac Lab, train a better policy, and then use it on the same H2 Plus robot that Stanford used last semester. This makes experiments repeatable, so the field can build on past work instead of starting over each time.
NVIDIA Isaac GR00T Reference Humanoid Robot Specs, Cost, and Availability
For research administrators reviewing budgets, the NVIDIA Isaac GR00T reference humanoid robot specs cost discussion begins at $29,900 the listed price for the H2 Plus-based system. The H2 Plus is expected to ship from Unitree in October 2026. Researchers who want to get started sooner can use the Isaac GR00T reference workflow for the smaller, more common Unitree G1 on GitHub and Hugging Face before the full hardware is released.
For university labs focused on leading-edge humanoid physical AI, NVIDIA’s reference design offers the most direct route from buying equipment to starting policy research that the field has seen to date. However, U.S. lawmakers have recently introduced the bipartisan American Security Robotics Act, a proposed bill that would ban federal purchases of Chinese-made unmanned ground vehicles due to concerns about data security and national security. Federally funded programs should watch this legislation closely before making any purchases.
The Structural Shift This Platform Represents
Closed robotics systems have forced every lab to pay a hidden cost: reconstructing infrastructure from the ground up. NVIDIA’s approach is to eliminate that cost through a standardized, open reference design, so researchers can focus on the real challenges such as locomotion recovery, dexterous manipulation, and language-based task execution.
Michael Yip, professor at UC San Diego and director of the Sophisticated Robotics and Controls Laboratory, noted that “an integrated platform that connects robot hardware, data capture, policy learning, and physical evaluation can help researchers accelerate loco-manipulation research and develop more useful real-world systems. “There hospitals and businesses will see walking, capable physical AI systems this decade depends on how quickly this progress happens. The NVIDIA Isaac GR00T Reference Humanoid Robot has given researchers the strongest starting point the field has ever had.
Source: NVIDIA Announces NVIDIA Isaac GR00T Reference Humanoid Robot for Academic Research
