NVIDIA Isaac GR00T is an open reference platform for general-purpose humanoid robots. It bundles open robot foundation models, open data and data pipelines, simulation frameworks built on Omniverse and Cosmos, middleware, CUDA-X runtime libraries and Jetson Thor hardware for real-time inference and control. The foundation models take language and images as input and perform manipulation tasks across different robot bodies. They are trained on real captured data, synthetic data and internet-scale video, and the current release is Isaac GR00T 1.7, available from Hugging Face.
The audience is technical: robotics engineers, machine learning researchers, humanoid startups and corporate R&D teams who need a head start on robot intelligence. Developers who already work with NVIDIA tools such as Isaac Lab and Isaac Sim will find a continuous path from training to validation to deployment. Academic labs can post-train the open models for their own robot embodiment, while enterprise teams evaluate them for industrial tasks. It assumes comfort with robot learning, simulation and GPU infrastructure rather than offering a no-code interface.
Models handle common tasks such as grasping, moving objects with one or both arms and handing items between arms, and can chain multi-step tasks that need long context. Documented application areas include material handling, packaging and inspection. Isaac-Teleop collects human demonstrations through teleoperation in the real world and in simulation, and GR00T-WholeBodyControl supplies libraries, models and policies for responsive whole-body control. A typical workflow trains policies in Isaac Lab, validates them in Isaac Sim, trains at scale on DGX Cloud, and deploys the result on Jetson AGX Thor inside the robot.
