NVIDIA and Hugging Face Open Up Humanoid Robot Brains to 19 Million Developers

NVIDIA and Hugging Face announced a deepened collaboration this week that pushes some of the most advanced tools in physical AI into the open-source robotics library LeRobot. The partnership brings NVIDIA Isaac GR00T 1.7 — a reasoning vision-language-action (VLA) model for humanoid robots — and the NVIDIA Isaac Teleop data-collection framework directly into LeRobot's workflows, with NVIDIA Cosmos 3, a frontier world model for physical AI, slated to follow soon. The announcement, reported by The Robot Report and NVIDIA's own engineering blog, effectively links NVIDIA's roughly 3 million robotics developers with Hugging Face's 16 million AI builders under one shared toolchain.

For an industry that has spent years fragmented across proprietary datasets, incompatible simulators, and closed foundation models, this is a notable structural shift. Instead of every robotics team building its own data pipeline from scratch, developers can now pull a benchmarked, commercially viable foundation model, fine-tune it on their own robot's morphology, and validate it — all inside one open ecosystem.

What's Actually New: GR00T 1.7, Isaac Teleop, and Cosmos 3

Isaac GR00T 1.7 is described as the first open and commercially viable robot foundation model, meaning companies can legally deploy it in products rather than treating it purely as a research artifact. It's a VLA model — it takes in camera video and natural-language instructions and outputs motor commands — and developers can post-train it onto new robot bodies and tasks using LeRobot's existing workflows. Isaac Teleop, meanwhile, solves a much less glamorous but critical problem: collecting clean human demonstration data. It standardizes how external devices capture teleoperated movements, so that datasets recorded on different hardware can be pooled and shared across the community rather than staying locked in isolated formats. The still-unreleased Cosmos 3 is positioned as a "world model" — a system that can simulate physics and generate synthetic training scenarios when real-world data collection is too slow or expensive.

The Numbers Behind the Hype

LeRobot already anchors a sizeable slice of the open robotics stack. Hugging Face says its open dataset collection tied to the platform has been downloaded more than 15 million times and includes over 350,000 real and simulated robot trajectories along with 57 million recorded grasps — a dataset scale that would have been unthinkable for an open project just a few years ago. The new integrations layer onto existing NVIDIA infrastructure already wired into LeRobot, including Isaac Sim and Isaac Lab simulation frameworks, the Isaac Lab-Arena environment hub for training generalist policies like GR00T, Pi, and SmolVLA, and a Jetson Thor hardware integration that lets developers deploy VLA models directly onto Hugging Face's open-source Reachy 2 humanoid.

"Open source is how a field turns advanced research into something people can study, adapt and build on. 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."

Thomas Wolf, co-founder and Chief Science Officer, Hugging Face

That quote gets at the strategic logic here: NVIDIA doesn't sell robots, it sells the compute and software stack underneath them, so widening the pool of developers building on Isaac tools strengthens its position regardless of which humanoid company ultimately wins. Hugging Face, in turn, cements LeRobot as the default meeting point for that activity — a repeat of the playbook that made its Transformers library the default hub for language models.

Quick Q&A: What This Means for Robotics Builders

Q: Is Isaac GR00T 1.7 free to use commercially, or just for research?

A: NVIDIA and Hugging Face describe it as the first open and commercially viable robot foundation model, meaning it's licensed for real product deployment, not restricted to academic use.

Q: What problem does Isaac Teleop actually solve?

A: It standardizes how teleoperated demonstration data is captured across different external devices, so datasets from different labs and companies can be pooled and reused instead of staying siloed in incompatible formats.

Q: Why does NVIDIA Cosmos 3 matter if it isn't released yet?

A: Cosmos 3 is meant to generate synthetic training data and simulate scenarios, addressing the industry's biggest bottleneck — the sheer cost and slowness of collecting enough real-world robot data.

Q: Which robots can already use these tools?

A: NVIDIA's Jetson Thor integration with LeRobot's Reachy 2 humanoid already supports deploying VLA models, and generalist policies like GR00T, Pi, and SmolVLA can be trained and evaluated through the Isaac Lab-Arena environment hub.

The move lands amid a broader wave of physical AI announcements this week — from Nomagic's vision-language-action deployment in live warehouses to Mistral AI's release of a navigation model trained entirely in simulation — suggesting the industry is converging on a shared belief: general-purpose "robot brains" trained on shared data will outcompete narrowly hand-coded systems. Whether GR00T, Cosmos, and their open-source rivals can close the gap between demo-day choreography and messy real-world deployment is the question the next year of shipments, not announcements, will actually answer.

Source: The Robot Report — "NVIDIA and Hugging Face bring new models and frameworks to LeRobot" (also reported on the NVIDIA Blog)

Published: July 8, 2026