NVIDIA Unveils Isaac GR00T Humanoid Reference Platform

NVIDIA announced the Isaac GR00T Reference Humanoid Robot, described as the first open humanoid reference design for academic research, at NVIDIA GTC Taipei, according to NVIDIA's press release. The system pairs a Unitree H2 Plus chassis and dual Sharpa Wave tactile five-finger hands with NVIDIA Jetson AGX Thor onboard compute and the open Isaac GR00T software stack and foundation models. NVIDIA frames it as a single integrated platform spanning data capture, simulation, training, evaluation, and deployment, intended to cut the fragmented integration work that slows humanoid research. Named research adopters include Ai2, ETH Zurich, Stanford Robotics Center, and UC San Diego's Advanced Robotics and Controls Laboratory.

Per NVIDIA, the robot stands nearly six feet tall and weighs about 150 pounds, with 75 total degrees of freedom (31 in the body plus 22 per Sharpa Wave hand). It reports arm torque up to 120 Newton-meters, leg torque up to 360 Newton-meters, a rated arm payload of 7 kilograms and peak payload of 15 kilograms, multi-view stereo and wrist cameras, and an on-remote emergency stop. Onboard compute is the NVIDIA Jetson AGX Thor T5000, with a Blackwell GPU rated at 2,070 FP4 teraflops, a 14-core Arm CPU, and 128GB of unified memory.

The Isaac GR00T platform includes Isaac Teleop for demonstration capture, open GR00T foundation models for reasoning and multitask behavior, Isaac Sim and Isaac Lab for simulation and training, accelerated Isaac ROS middleware for deployment, and Jetson Thor for on-robot inference. The GR00T N1 generation of the foundation model is documented on arXiv as a Vision-Language-Action dual-system architecture, with a vision-language module and a diffusion-transformer motor module trained on a mix of real robot trajectories, human video, and synthetic data (arXiv:2503.14734).

Offering a packaged hardware-and-software reference, rather than ad hoc lab builds, can let research teams move faster from robot bring-up to skill development. The open artifacts and the public model paper increase transparency relative to closed robot stacks and may aid reproducibility. CNBC and other outlets note NVIDIA selected Unitree, a Chinese robotics company, for the first publicly available system, tying the announcement to commercial humanoid suppliers and to questions about hardware availability.

• •Follow-up benchmarks and reproducibility studies from the named research labs.
• •Published sim-to-real transfer results and failure-mode analyses on diverse embodiments.
• •Additional hardware pairings, since NVIDIA says the platform will also support the Unitree G1.
• •Availability, which NVIDIA says will be from Unitree in late 2026, and any export-control or supply constraints.