NVIDIA Sets an Open Base for Humanoids

NVIDIA announced on May 31, 2026, the Isaac GR00T Reference Humanoid Robot, an open reference design for humanoid robotics research. The system combines a Unitree H2 Plus body, Sharpa Wave tactile five-finger hands, Jetson Thor onboard compute and the Isaac GR00T software platform. According to the official announcement, it will be available from Unitree in late 2026. The verified point is narrow but significant: NVIDIA is not only presenting an AI model for robots, but a full hardware and software configuration meant for labs that want to test humanoid behaviors on a real machine.

The word "reference" matters. In robotics, many teams spend a large share of their time stitching together separate components: the mechanical body, hands, cameras, teleoperation tools, simulation, learning systems and deployment software. NVIDIA is proposing a shared base. The release describes a humanoid standing nearly six feet tall, weighing 150 pounds, with 31 degrees of freedom across the body and 75 degrees of freedom when the hands are included. It also lists a head-mounted wide-angle stereo camera, wrist cameras, an inertial measurement unit and around three hours of battery life. Degrees of freedom are the controllable axes of motion, such as an arm joint or finger joint.

The research problem is less flashy than the humanoid image suggests, but it is central. General-purpose humanoids do not advance only through larger models. They need demonstration data, simulated scenarios, reproducible tests and a cleaner path from lab experiments to real robots. The Isaac GR00T stack is meant to cover that pipeline: Isaac Teleop for capturing human demonstrations, Isaac Sim and Isaac Lab for simulation and policy evaluation, open foundation models for humanoid reasoning and multitask behavior, and Isaac ROS for moving trained policies onto the robot. A control policy is the software that chooses a robot’s next action from sensor input and a goal.

What changes in practical terms is the chance for several research groups to compare methods on a common platform. NVIDIA names Ai2, ETH Zurich, the Stanford Robotics Center and UC San Diego’s Advanced Robotics and Controls Laboratory among the institutions that will use the design. Caution is still needed: the announcement includes features that depend on future availability, and a reference robot does not by itself solve safety, cost or reliability in real environments. But it marks a useful shift for humanoid robotics: moving part of the competition away from basic integration and toward data, learning methods and the tasks these machines can actually perform.