GR00T 1.7 and Isaac Teleop give LeRobot an open humanoid pipeline

Robotics has not had the same open flywheel as language models because data collection, simulation, validation, and deployment are still fragmented and expensive. NVIDIA and Hugging Face are trying to narrow that gap by bringing Isaac GR00T 1.7, Isaac Teleop, datasets, and workflows into LeRobot, Hugging Face’s open source robotics library.

LeRobot is built for training, running, and sharing robot datasets, models, policies, and workflows. NVIDIA says the integration links its 3 million robotics developers with Hugging Face’s 16 million AI builders. The goal is a common pipeline: collect and standardize robot data, post-train robot foundation models, evaluate policies in simulation, then deploy through open workflows.

The pieces are practical rather than only aspirational. Isaac Teleop gives developers an open source framework for capturing human demonstrations from external devices in interoperable formats. Isaac GR00T 1.7 brings an open, commercially viable vision-language-action model for humanoid robots into LeRobot workflows, so teams can adapt it to new embodiments and tasks. NVIDIA also plans to bring Cosmos 3, a world foundation model for physical AI, into the same loop for data generation, simulation, and policy development when real-world data is scarce.

The data scale gives the integration weight. NVIDIA points to an open source physical AI dataset downloaded more than 15 million times, with more than 350,000 real and simulated trajectories and 57 million grasps. It also ties in Isaac Sim, Isaac Lab, Isaac Lab-Arena, and Jetson Thor support for Reachy 2. For humanoid and physical AI builders, the useful shift is not one model name; it is a more reproducible path from demonstrations to training to simulated checks to robot deployment.