How can robots learn complex tasks? An open approach brings AI models, datasets and simulation tools into one development workflow.
NVIDIA has expanded its collaboration with Hugging Face to integrate its robotics AI stack into the LeRobot open-source library, giving robotics developers access to robot foundation models, teleoperation tools and simulation frameworks through a unified development workflow.
The collaboration aims to address one of the biggest engineering challenges in physical AI, building and validating robots requires large-scale datasets, simulation environments and specialised AI models that are often fragmented across different platforms. By bringing these technologies into LeRobot, developers can collect training data, fine-tune robot foundation models, evaluate policies and deploy robots using a common open-source framework.
Among the new additions is NVIDIA Isaac GR00T 1.7, an open vision-language-action (VLA) foundation model designed for humanoid robots. The model enables post-training for new robot embodiments and task-specific applications without developing models from scratch. NVIDIA has also integrated Isaac Teleop, an open-source framework that captures human demonstrations from external input devices in standardised formats for imitation learning and dataset creation.
The roadmap also includes support for NVIDIA Cosmos 3, a world foundation model that can generate synthetic robotics data, simulate real-world scenarios and assist policy development when collecting physical data is impractical or expensive. Existing integrations already available through LeRobot include Isaac Sim and Isaac Lab simulation frameworks, Isaac Lab-Arena for creating and benchmarking simulation environments, a physical AI dataset containing more than 350,000 real and simulated robot trajectories and 57 million grasp samples, and deployment support for vision-language-action models on Jetson Thor-powered humanoid robots.
“Open source is how a field turns advanced research into something people can study, adapt and build on,” says Thomas Wolf, Co-founder and Chief Science Officer at Hugging Face. “With NVIDIA Isaac GR00T 1.7 and Isaac TeleOp in LeRobot today, robotics developers can use shared models, data and workflows to train and evaluate robots in the open.”
