Nvidia has released 100,000 hours of robotics training data and a set of development tools to support the use of its GR00T 2 foundation model. The company made its hardware designs, agentic framework, and simulation and annotation engines available to the public today to facilitate the creation of autonomous systems. While the supporting tools are now accessible, Nvidia will not release the GR00T 2 model weights and X-Adapt recipes until December.
The GR00T 2 system runs entirely on a robot’s onboard compute to handle long-horizon tasks, such as assembling model airplanes, that involve more than 20 steps and last over five minutes. The model is designed to operate across different physical platforms, from parallel grippers to humanoids, and utilizes a process called X-Adapt to deploy data from simulations and human wearables directly to real robots.
