NVIDIA, Hugging Face integrate Isaac GR00T and Cosmos 3 into open LeRobot library

NVIDIA and Hugging Face have integrated NVIDIA Isaac GR00T 1.7, an open vision-language-action (VLA) model for humanoid robots, and the NVIDIA Isaac Teleop data-collection framework directly into LeRobot, Hugging Face's open-source robotics library, with NVIDIA Cosmos 3 (a frontier world model for physical AI) planned soon. The integrations give developers standardized workflows for end-to-end robot training, evaluation, and deployment; they connect NVIDIA's 3 million robotics developers with Hugging Face's 16 million AI builders, expanding access to frontier physical AI tools through open development loops.

Isaac GR00T 1.7 is the first open, commercially viable robot foundation model; developers can post-train and deploy it through LeRobot workflows, adapting it to new robot embodiments and tasks with benchmarked performance. Isaac Teleop provides a standardized framework for collecting high-quality human demonstrations from external devices (tele-operated robots), capturing data in interoperable formats directly within LeRobot so teams can expand and share datasets with the community. These tools integrate with NVIDIA's existing robotics infrastructure already connected to LeRobot: Isaac Sim and Isaac Lab simulation frameworks, the largest open-source physical AI dataset (350k+ real and simulated trajectories, 57m grasps), and Isaac Lab-Arena for rapid prototyping of complex simulation environments.

NVIDIA also announced Jetson Thor integration with LeRobot's Reachy 2 humanoid robot, enabling on-device VLA model deployment on edge hardware. Cosmos 3, arriving soon in LeRobot, will help developers generate synthetic robotics data, simulate scenarios, and support policy development when real-world data is scarce or expensive. This ecosystem positions developers to train and evaluate generalist robot policies (like GR00T) within a shared, version-controlled development loop.

For robotics teams and hardware builders, this represents a shift toward standardized, collaborative physical AI development. Previously fragmented resources (datasets, foundation models, simulation, compute validation) are now integrated in one platform. Open models + standardized deployment reduce lock-in and enable faster iteration. For NVIDIA, it deepens platform stickiness: Jetson hardware becomes the de-facto standard for edge inference, while simulation on Isaac tightens the training loop. For enterprises building embodied AI systems, standardized LeRobot workflows lower the barrier to experiment with foundation models before committing to proprietary toolchains.