What the Isaac GR00T Humanoid Reference Design Actually Is
The Isaac GR00T humanoid reference design is an open, human-scale robotics research platform that combines a full humanoid body, onboard Jetson Thor compute, and a complete Isaac GR00T software stack so labs can build, train, and test practical robot skills without assembling fragmented hardware and code on their own. Announced as the NVIDIA Isaac GR00T Reference Humanoid Robot, it is built around the Unitree H2 Plus chassis, Sharpa Wave tactile five-finger hands, and NVIDIA Jetson AGX Thor T5000 compute module. Jensen Huang introduced the package during NVIDIA’s GTC event in Taipei on May 31 as a single starting point for academic and frontier robotics research. Instead of stitching together bodies, controllers, and simulators from different vendors, teams gain a unified humanoid robot development pipeline that spans simulation, training, and real-world deployment on the same physical robot.
Hardware: Human-Scale Body with Tactile, High-DOF Hands
At the heart of the Isaac GR00T humanoid reference design is the Unitree H2 Plus, a human-scale platform standing around six feet tall and weighing about 150 pounds. The chassis delivers 31 degrees of freedom across the body, enabling human-like walking, bending, and reaching for realistic physical AI experiments. NVIDIA’s configuration adds dual Sharpa Wave five-finger hands, which raise the system to 75 total degrees of freedom and give it tactile sensing for delicate, dexterous manipulation. Multi-view sensing includes a head-mounted stereo camera with a 140-degree horizontal and 102-degree vertical field of view, wrist cameras for close-up manipulation tasks, and an inertial measurement unit for motion tracking. The legs can generate up to 360 Newton-metres of torque and the arms up to 120, with each arm rated for a 7kg payload and a 15kg peak payload, making the robot suitable for demanding lifting and reach experiments.
Jetson Thor Compute and the Isaac GR00T Software Stack
The Isaac GR00T humanoid reference design is as much a computing and software platform as it is a robot body. Onboard compute comes from the NVIDIA Jetson AGX Thor T5000 module, which integrates a Blackwell GPU delivering 2,070 FP4 teraflops of AI performance, a 14-core Arm CPU, and 128GB of unified memory within a configurable 40W–130W power envelope. According to NVIDIA, this combination gives researchers “a single, open platform to make breakthrough discoveries toward general-purpose physical intelligence.” Connectivity includes Ethernet, Wi‑Fi 6, Bluetooth 5.2, USB, microphones, and speakers, plus a remote emergency stop, while a 15Ah 0.972kWh battery supports several hours of operation. On the software side, Isaac GR00T provides open models and workflows for perception, control, and policy learning, enabling teams to move from simulated tasks to real-world execution without rewriting core logic for different robots or stacks.
SharpaWave Hands and Unitree H2 Plus: Core Partners in the Ecosystem
Two key hardware partnerships illustrate how NVIDIA is building an ecosystem around the Isaac GR00T humanoid platform. SharpaWave supplies the tactile five-finger hands that serve as the robot’s primary manipulators, bringing 22 degrees of freedom to the hands alone and enabling careful object handling that simple grippers cannot match. These hands, combined with multi-view sensing, allow experiments in in-hand manipulation, tool use, and contact-rich tasks that are essential for practical humanoid robot development. Unitree, meanwhile, provides the H2 Plus humanoid chassis that forms the physical base of the reference design. The H2 Plus reference robot, built on Isaac GR00T, combines this chassis with Jetson Thor compute and SharpaWave hands into a single package targeted at frontier humanoid research. With NVIDIA’s open software stack “at the center,” Unitree positions H2 Plus as a state-of-the-art platform for physical AI development rather than a closed, proprietary product.
Democratizing Humanoid Robot Development for Labs and Startups
By releasing Isaac GR00T as an open reference design, NVIDIA aims to turn humanoid robots into a shared robotics research platform rather than a bespoke engineering project for each lab. Academic teams and startups gain a pre-integrated system that spans hardware integration, sensing, whole-body control, and AI workflows, so they can shift focus toward skill learning, evaluation, and real-world testing. Unitree’s H2 Plus, built on Isaac GR00T, shows early ecosystem adoption, giving developers a ready-made humanoid with tactile hands and Jetson Thor compute. Research groups can standardize on this platform, share code, and reproduce experiments more easily, accelerating progress toward practical humanoid behaviors. Instead of spending months on bring-up and calibration, labs can move faster into problems like safe human-robot interaction, industrial assistance, and household tasks, which is exactly the kind of multitrillion-dollar opportunity Jensen Huang describes for “general-purpose physical intelligence.”
