NVIDIA's open robot model lands in LeRobot
Isaac GR00T 1.7, billed as the first open, commercially viable robot foundation model, and the Isaac Teleop data tool are now integrated into LeRobot.
NVIDIA and Hugging Face integrated Isaac GR00T 1.7, which NVIDIA describes as the first open, commercially viable robot foundation model, along with the Isaac Teleop data-collection framework into LeRobot, Hugging Face’s open-source robotics library, on July 7. NVIDIA also previewed a forthcoming Cosmos 3 world-model integration for the same library.
GR00T 1.7 is a reasoning vision-language-action model, meaning it can take a camera feed and a natural-language instruction and reason about the steps needed before acting, rather than mapping inputs to actions directly. The release lands alongside LeRobot’s own v0.6.0 update, which adds five new vision-language-action models including GR00T N1.7 and MolmoAct2, plus new world-model policies and six simulation benchmarks unified under one evaluation command.
NVIDIA frames the integration as connecting its roughly 3 million registered robotics developers with Hugging Face’s 16 million builders. LeRobot’s physical-AI dataset now has more than 15 million downloads, 350,000-plus recorded trajectories, and 57 million grasps, a meaningful open corpus for anyone training or fine-tuning robot-control policies without starting from scratch.
For teams working on physical AI, this is the practical value: a commercially usable foundation model plus a data-collection tool plus a benchmark suite, all in one open-source library, rather than stitching together research code from separate repos.