NVIDIA and Hugging Face Expand LeRobot With New AI Models and Frameworks for Open Robotics
Open-source AI has transformed software development by making powerful models, datasets, and tools accessible to a global community of developers. Robotics is now following a similar path, with open platforms helping accelerate innovation in physical AI. However, building intelligent robots still requires access to large datasets, advanced robot foundation models, simulation environments, and powerful computing resources—tools that have traditionally been expensive or difficult to obtain.
To make robotics development more accessible, NVIDIA and Hugging Face have announced new integrations for LeRobot, Hugging Face’s open-source robotics framework. The collaboration brings NVIDIA Isaac GR00T 1.7, the company’s open vision-language-action (VLA) model for humanoid robots, and the NVIDIA Isaac Teleop framework into LeRobot. NVIDIA also plans to integrate Cosmos 3, its next-generation world foundation model for physical AI, enabling developers to create richer training data and simulations.
These additions provide robotics developers with an open, standardized workflow for collecting data, training robot foundation models, evaluating performance, and deploying intelligent robots. By combining NVIDIA’s robotics technologies with Hugging Face’s collaborative AI ecosystem, the partnership aims to lower barriers to robotics research while encouraging greater innovation across the global open-source community.
Thomas Wolf, co-founder and Chief Science Officer of Hugging Face, emphasized that open-source software has historically helped transform cutting-edge research into practical technologies that developers can study, improve, and build upon. He noted that bringing NVIDIA Isaac GR00T 1.7 and Isaac Teleop into LeRobot gives robotics developers access to shared models, datasets, and workflows for training and evaluating robots. The planned integration of NVIDIA Cosmos 3 will further enable the community to incorporate advanced world models into open robotics development.
LeRobot serves as an open-source robotics library that allows developers to train, test, share, and deploy robot datasets, models, and policies. NVIDIA’s continued collaboration with Hugging Face connects millions of robotics developers with one of the world’s largest AI communities, expanding access to advanced physical AI technologies through open development practices.
One of the key additions is NVIDIA Isaac Teleop, an open-source framework designed to simplify robot data collection. Developers can capture high-quality demonstrations of human actions using external devices, convert the information into standardized formats, and easily expand or share datasets with other researchers through LeRobot. This helps create larger, more consistent datasets that improve robot learning.
Another major integration is NVIDIA Isaac GR00T 1.7, the first commercially viable open robot foundation model. Through LeRobot workflows, developers can fine-tune the model for different robot designs and applications, making it easier to adapt robots for new environments and tasks while maintaining strong benchmark performance.
Looking ahead, NVIDIA plans to integrate Cosmos 3 into LeRobot. The upcoming frontier world foundation model will help developers generate synthetic robotics data, augment existing datasets, simulate complex scenarios, and support policy development when collecting real-world data is too expensive or impractical. This capability is expected to significantly reduce development costs while improving the quality and diversity of training data.
The new integrations complement several NVIDIA resources already available through LeRobot. Developers can access one of the world’s largest open-source physical AI datasets, featuring more than 350,000 real and simulated robot trajectories along with over 57 million grasp samples. The dataset has already been downloaded more than 15 million times, demonstrating its value to the robotics research community.
LeRobot also supports NVIDIA Isaac Sim and Isaac Lab, enabling developers to build simulation environments, generate training data, evaluate robot behaviors, and validate policies before deploying them on physical robots. These simulation tools reduce development time while improving reliability.
Another addition is NVIDIA Isaac Lab-Arena, now available through the LeRobot Environment Hub. Developers can rapidly prototype complex simulation environments, register them within LeRobot, and use them to train and evaluate general-purpose robot policies, including GR00T, Pi, and SmolVLA models.
The collaboration also extends to deployment hardware through NVIDIA Jetson Thor integration with Reachy 2, allowing developers to run advanced vision-language-action models on open-source humanoid robots with greater efficiency.
Together, these integrations establish LeRobot as a comprehensive open robotics platform that supports the entire development lifecycle—from data collection and simulation to model training, evaluation, and real-world deployment. By combining NVIDIA’s physical AI technologies with Hugging Face’s open-source ecosystem, developers gain access to powerful tools that accelerate robotics innovation while fostering collaboration across the global research community.
As robotics continues to evolve, open platforms such as LeRobot are expected to play an increasingly important role in making advanced physical AI technologies accessible to researchers, startups, and developers worldwide, helping drive the next generation of intelligent robots.
Source link: https://blogs.nvidia.com