NVIDIA and Hugging Face expand LeRobot with Isaac models and planned Cosmos 3

NVIDIA said it is bringing new pieces of its “physical AI” stack into LeRobot, the open robotics library maintained by Hugging Face, in an effort to make robot training and deployment workflows more accessible to developers. In a blog post published Monday, the companies said the integration includes NVIDIA’s Isaac GR00T 1.7 model for humanoid robots and NVIDIA’s Isaac TeleOp control framework, with a next step planned to connect NVIDIA Cosmos 3, a frontier model for physical AI, into the same open development loop.

At the center of the announcement is LeRobot itself, described by NVIDIA as an open source library for training, running, and sharing robot datasets, models, policies, and workflows. By connecting NVIDIA’s robotics tooling to LeRobot, the companies argue that developers can build a standardized pathway from data collection to evaluation and deployment, rather than stitching together separate tools and model ecosystems.

NVIDIA and Hugging Face said Isaac GR00T 1.7 will be incorporated into LeRobot. The blog describes GR00T 1.7 as an “open, reasoning vision language action (VLA) model” for humanoid robots. In plain terms, a vision-language-action model takes in visual inputs and language instructions, then produces action outputs, and the “reasoning” framing suggests the model is meant to follow through on intermediate decision-making steps for physical tasks.

The companies also said Isaac TeleOp, a robotics teleoperation framework, is being brought to LeRobot. Teleoperation typically means humans can guide a robot in real time or record demonstrations, creating training indicates that later help models learn. NVIDIA’s post positions the TeleOp integration as part of a broader effort to streamline data collection and the training/evaluation cycle for humanoids.

NVIDIA further tied the integration effort to its broader physical AI roadmap by pointing to Cosmos 3, which it described as “planned soon.” Cosmos 3 is characterized as a frontier model for physical AI, which the post frames as the next ingredient developers will want to plug into open workflows. Hugging Face cofounder and chief science officer Thomas Wolf said open-source collaboration is how research becomes adaptable technology, and he argued that having shared models and workflows in LeRobot lets developers train and evaluate robots “in the open,” with a route to fold frontier “world models” into that loop later.

NVIDIA’s post also highlighted the scale of each community involved. It said NVIDIA has around 3 million robotics developers, while Hugging Face has about 16 million AI builders, describing the partnership as a way to broaden access to physical AI tools through open workflows. The company did not provide revenue or customer numbers tied to the LeRobot integrations in the announcement.

The blog positioned the technical value proposition as a reduction in fragmentation across the end-to-end robot development lifecycle. It said bringing NVIDIA physical AI capabilities into LeRobot gives developers “a common way” to collect and standardize data, train and fine-tune robot foundation models, evaluate performance, and deploy models through open workflows. “Robot foundation models” are generally large, broadly capable models that can be adapted to specific tasks via fine-tuning, which the companies are implicitly using as a benchmark for what they want developers to access more easily.

Beyond Monday’s specific integration, NVIDIA pointed readers to a broader set of resources it says are already connected to LeRobot, including Isaac Lab-Arena as part of the overall humanoid development pathway. NVIDIA said the combined resources are intended to support end-to-end development, from collection and training to evaluation in a standardized environment, and it referred to “detailed” LeRobot integration workflows without outlining all the specifics in the post itself.

What is not fully spelled out in the announcement is exactly what changes developers will need to make in their existing LeRobot setups to adopt the new Isaac components, including compatibility requirements, rollout timelines for any “planned” items, or the level of performance benchmarking that the companies expect from the integrations. The post also does not provide a date for Cosmos 3’s LeRobot integration beyond “planned soon,” nor does it specify how model licensing, access controls, or hardware prerequisites will work for all users.

Developers and industry watchers may want to track when Cosmos 3 support lands in LeRobot, as well as any accompanying documentation that clarifies evaluation methods and dataset standards for humanoid robots. The practical test for the announcement will be whether open community adoption increases, because the companies are betting that a shared workflow and common model entry points can accelerate experimentation and reduce the cost of building physical AI systems.