It is not a consumer product. It is a research instrument: a purchasable, repeatable blueprint that pairs a full-size humanoid body with high-fidelity hands, a powerful onboard AI brain, and an open software stack, so labs can spend less time building robots and more time training them.

A Single Platform: Body, Hands, and Brain

The Isaac GR00T reference design bundles three hardware layers into one integrated system :

• Unitree H2 Plus chassis – The robot stands nearly 6 feet tall (≈180 cm) and weighs about 150 pounds (68 kg). Its body provides 31 degrees of freedom (DOF) with a maximum arm torque of 120 N·m and leg torque of 360 N·m . A 0.972 kWh battery keeps it running for roughly three hours, and head-mounted stereo cameras, wrist cameras, array microphones, and speakers handle perception and voice interaction .
• Sharpa Wave five-fingered hands – Each hand adds 22 degrees of freedom, bringing the total system to 75 DOF . Weighing just 1.3 kg each, these tactile, dexterous end-effectors are designed for precise tasks that stump most humanoids: assembling a computer, handling surgical tools, or inserting an IV .
• NVIDIA Jetson Thor onboard compute – At the center sits the Jetson AGX Thor module with a Blackwell-architecture GPU, delivering up to 2,070 TFLOPS of FP4 AI performance . It runs the full Isaac GR00T software stack—simulation, model training, policy deployment—directly on the robot .

Because the design is fully open, researchers can modify the hardware, train custom policies in NVIDIA Isaac Sim, and deploy them on the physical unit without being locked to a single vendor’s ecosystem .

Why a Reference Design Matters

The term “reference design” is borrowed from the smartphone industry. Just as Google’s Android reference phones gave manufacturers a known-good starting point while letting them innovate on top, the GR00T platform aims to give robotics labs a common baseline .

The goal is to accelerate sim-to-real transfer: training manipulation policies in high-fidelity simulation, then running them on identical physical hardware. Today, a walking controller developed in one lab rarely works on another’s robot. NVIDIA’s wager is that a standardized, open platform will change that .

The robot will be available to researchers in late 2026, with some sources pointing to an October 2026 release window .

Five Labs Already Onboard

Five prominent institutions have already committed to using the platform :

• Stanford University
• ETH Zurich
• Allen Institute for AI (AI2)
• UC San Diego – Advanced Robotics and Controls Laboratory
• University of Tokyo – JSK Robotics Laboratory

These early adopters span the US, Europe, and Asia, giving the platform immediate geographic and research diversity.

The Geopolitical Dimension

A notable subtext is the partnership itself. A US chip giant (NVIDIA) is collaborating with a leading Chinese hardware manufacturer (Unitree) and a Singaporean component firm (Sharpa) in a domain—advanced robotics—that sits squarely at the intersection of commercial ambition and national security concern . The cross-border nature of the project was a significant talking point in early coverage .

Agentic AI and the Bigger Computex Picture

The GR00T robot was part of a much broader keynote. Huang used the Computex stage to declare 2026 the year of “agentic AI”—autonomous, goal-driven AI systems operating across the entire computing stack .

Beyond the robot, the two other major announcements from the keynote were :

• NVIDIA RTX Spark – A new Arm-based superchip (developed with MediaTek) aimed at “reinventing the PC” for local AI agents, with up to 20 CPU cores and 6,144 CUDA cores.
• Vera Rubin AI infrastructure – NVIDIA’s next-generation data-center platform, now in production, designed specifically for agentic AI workloads at scale.

In this context, the Isaac GR00T robot is the physical embodiment of the agentic AI thesis: if software agents are going to reason and act autonomously, the hardware that lets them manipulate the real world needs a common development platform. NVIDIA just shipped its first one.

It is not a consumer product. It is a research instrument: a purchasable, repeatable blueprint that pairs a full-size humanoid body with high-fidelity hands, a powerful onboard AI brain, and an open software stack, so labs can spend less time building robots and more time training them.

A Single Platform: Body, Hands, and Brain

The Isaac GR00T reference design bundles three hardware layers into one integrated system :

• Unitree H2 Plus chassis – The robot stands nearly 6 feet tall (≈180 cm) and weighs about 150 pounds (68 kg). Its body provides 31 degrees of freedom (DOF) with a maximum arm torque of 120 N·m and leg torque of 360 N·m . A 0.972 kWh battery keeps it running for roughly three hours, and head-mounted stereo cameras, wrist cameras, array microphones, and speakers handle perception and voice interaction .
• Sharpa Wave five-fingered hands – Each hand adds 22 degrees of freedom, bringing the total system to 75 DOF . Weighing just 1.3 kg each, these tactile, dexterous end-effectors are designed for precise tasks that stump most humanoids: assembling a computer, handling surgical tools, or inserting an IV .
• NVIDIA Jetson Thor onboard compute – At the center sits the Jetson AGX Thor module with a Blackwell-architecture GPU, delivering up to 2,070 TFLOPS of FP4 AI performance . It runs the full Isaac GR00T software stack—simulation, model training, policy deployment—directly on the robot .

Because the design is fully open, researchers can modify the hardware, train custom policies in NVIDIA Isaac Sim, and deploy them on the physical unit without being locked to a single vendor’s ecosystem .

Why a Reference Design Matters

The term “reference design” is borrowed from the smartphone industry. Just as Google’s Android reference phones gave manufacturers a known-good starting point while letting them innovate on top, the GR00T platform aims to give robotics labs a common baseline .

The goal is to accelerate sim-to-real transfer: training manipulation policies in high-fidelity simulation, then running them on identical physical hardware. Today, a walking controller developed in one lab rarely works on another’s robot. NVIDIA’s wager is that a standardized, open platform will change that .

The robot will be available to researchers in late 2026, with some sources pointing to an October 2026 release window .

Five Labs Already Onboard

Five prominent institutions have already committed to using the platform :

• Stanford University
• ETH Zurich
• Allen Institute for AI (AI2)
• UC San Diego – Advanced Robotics and Controls Laboratory
• University of Tokyo – JSK Robotics Laboratory

These early adopters span the US, Europe, and Asia, giving the platform immediate geographic and research diversity.

The Geopolitical Dimension

A notable subtext is the partnership itself. A US chip giant (NVIDIA) is collaborating with a leading Chinese hardware manufacturer (Unitree) and a Singaporean component firm (Sharpa) in a domain—advanced robotics—that sits squarely at the intersection of commercial ambition and national security concern . The cross-border nature of the project was a significant talking point in early coverage .

Agentic AI and the Bigger Computex Picture

The GR00T robot was part of a much broader keynote. Huang used the Computex stage to declare 2026 the year of “agentic AI”—autonomous, goal-driven AI systems operating across the entire computing stack .

Beyond the robot, the two other major announcements from the keynote were :

• NVIDIA RTX Spark – A new Arm-based superchip (developed with MediaTek) aimed at “reinventing the PC” for local AI agents, with up to 20 CPU cores and 6,144 CUDA cores.
• Vera Rubin AI infrastructure – NVIDIA’s next-generation data-center platform, now in production, designed specifically for agentic AI workloads at scale.

In this context, the Isaac GR00T robot is the physical embodiment of the agentic AI thesis: if software agents are going to reason and act autonomously, the hardware that lets them manipulate the real world needs a common development platform. NVIDIA just shipped its first one.