NVIDIA turns GR00T into a reference robot

NVIDIA has announced a reference humanoid robot for academic research, built around its Isaac GR00T platform. The system combines a Unitree H2 Plus humanoid body, Sharpa Wave tactile five-finger hands, Jetson AGX Thor onboard compute and the open Isaac GR00T software stack. The central fact is not that NVIDIA is suddenly selling a general-purpose humanoid of its own. It is offering a full reference design, meant to give research labs a shared hardware and software base for testing robot behaviors in the real world.

A reference robot is a reproducible configuration: teams know which body, sensors, compute module and software tools are supposed to work together. In its announcement, NVIDIA describes a nearly six-foot humanoid weighing 150 pounds, with 31 degrees of freedom in the body, two hands with 22 degrees of freedom each, cameras in the head and wrists, an inertial measurement unit and a battery rated for about three hours. The Jetson AGX Thor T5000 provides onboard compute for sensor processing, reasoning and robot control, reducing dependence on a remote server for every decision.

The practical robotics angle is clear. Many humanoid projects are slowed down by fragmented integration: choosing a robot, connecting dexterous hands, capturing demonstrations, simulating motions, training control policies and then checking whether those policies still work on a physical machine. Isaac GR00T tries to connect those steps through Isaac Teleop for demonstration capture, open GR00T models for learning and reasoning, Isaac Sim and Isaac Lab for simulation, then Isaac ROS and Jetson Thor for deployment on the robot. That does not prove that a humanoid will perform useful work without supervision, but it lowers the entry cost for comparing research methods.

NVIDIA says Ai2, ETH Zurich, Stanford Robotics Center and UC San Diego’s Advanced Robotics and Controls Laboratory are among the institutions that will use the design. Availability is planned through Unitree in late 2026, while an Isaac GR00T workflow for the Unitree G1 is expected on GitHub and Hugging Face. Caution is still needed: the source is NVIDIA’s own announcement, and several pieces remain tied to delivery timelines. But the signal is meaningful. In humanoid robotics, the race is no longer only about demonstration videos. It is moving toward shared platforms where researchers can measure, reproduce and exchange robot behaviors with less initial integration work.