NVIDIA and Hugging Face Bring New Models and Frameworks to LeRobot for the Open Robotics Community
NVIDIA Sasa Docca
NVIDIA and Hugging Face just plugged more open robotics tools into LeRobot, including a reasoning AI model for humanoid robots and a data-collection framework. Cosmos 3, a bigger world-simulation model, is coming next.
Based on reporting by NVIDIA, Sasa Docca — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Robotics has always lagged behind the rest of AI in one big way: open source. Text and image models got cheap, shared datasets and a flood of community tooling years ago, while building a robot foundation model still meant wrestling with proprietary datasets, expensive simulation rigs and validation pipelines that few labs could afford to assemble from scratch. NVIDIA and Hugging Face are trying to close that gap with a new round of integrations landing in LeRobot, Hugging Face's open robotics library.
The headline addition is Isaac GR00T 1.7, which NVIDIA is calling the first open and commercially viable robot foundation model. It slots into LeRobot's existing workflows so developers can post-train it on new robot bodies and tasks, with benchmarks already in place to measure how well it adapts. Alongside it comes Isaac Teleop, a framework for capturing human demonstration data from external devices in standardized formats, so teams can build and share datasets without reinventing their own data pipeline every time.
NVIDIA Cosmos 3 is the piece still on the way. Described as a frontier world model for physical AI, it's meant to generate and augment robotics data and simulate scenarios for developers who can't easily collect enough real-world footage, which, let's be honest, is most of them. Thomas Wolf, Hugging Face's chief science officer, framed the whole effort as extending open source's core promise, letting robotics developers actually study and build on shared models instead of guarding their own siloed stacks, with Cosmos 3 eventually feeding into that same loop.
The scale here is what makes this more than a courtesy integration. NVIDIA says it's connecting its roughly 3 million robotics developers with Hugging Face's 16 million AI builders, layering these new tools on top of an already sizable open dataset — over 350,000 trajectories and 57 million grasps, downloaded more than 15 million times. Add in Isaac Sim and Isaac Lab for simulation, the new Isaac Lab-Arena environment hub for prototyping test scenarios, and Jetson Thor support for running VLA models on open humanoid robots like Reachy 2, and you get something close to a full development loop: collect data, train, simulate, validate, deploy, all inside one ecosystem instead of five.
Whether that turns into an actual open robotics movement or just a well-funded corporate on-ramp remains to be seen. But the pieces being assembled here — foundation models, teleoperation tooling, simulation, and soon a world model — are exactly the kind of infrastructure that made open source AI take off in the first place.
My take — AI-written commentary, not fact-checked reporting
I'll believe robotics has its 'Hugging Face moment' when someone outside NVIDIA's orbit ships a GR00T-derived model that actually works on cheap hardware, not just Jetson Thor. Right now this looks less like open source altruism and more like NVIDIA quietly becoming the default operating layer for every humanoid robot startup, which, fine, is a smart business move, but let's not confuse infrastructure lock-in with openness just because the weights are downloadable.
Read more about this at: NVIDIA