Skip to main content
Live
Main content

Nvidia and Hugging Face push Isaac GR00T 1.7 into LeRobot for open robotics

The integration connects 3M robotics developers to 16M AI builders, with Cosmos 3 world models coming next to Hugging Face's open library.

Jaeden Schafer
Editor in Chief · · 5 min read
Nvidia logo

Nvidia and Hugging Face are folding Nvidia's core physical AI stack into LeRobot, Hugging Face's open source robotics library, giving developers a single pipeline for collecting data, training foundation models and deploying policies on real hardware. The integration lands Isaac GR00T 1.7, the first open and commercially usable humanoid foundation model from Nvidia, alongside the Isaac Teleop data-collection framework directly inside LeRobot. Cosmos 3, Nvidia's frontier world model for physical AI, is planned to follow. The announced reach: 3 million robotics developers on Nvidia's side and 16 million AI builders on Hugging Face's.

The pitch is that robotics has been stuck where language models were before the open model wave — expensive datasets, siloed simulators, custom fine-tuning pipelines and no shared benchmark on which to compare policies. Nvidia is contributing what it already sells to industrial customers, and Hugging Face is contributing the distribution surface it built for open models.

Isaac GR00T 1.7 is a reasoning vision-language-action model aimed at humanoids. Inside LeRobot, developers can post-train it on new embodiments and tasks using standardized workflows, which is the piece that has been missing for teams that don't want to build their own training harness from scratch. Isaac Teleop handles the upstream problem — collecting high-quality human demonstrations from external devices in an interoperable format, then sharing the resulting datasets with the community.

Open source is how a field turns advanced research into something people can study, adapt and build on.
Thomas Wolf, Cofounder and Chief Science Officer, Hugging Face

Key facts

  • 01Nvidia is bringing Isaac GR00T 1.7 and Isaac Teleop to Hugging Face's LeRobot library, with Cosmos 3 world models planned next.
  • 02The partnership connects Nvidia's 3 million robotics developers with Hugging Face's 16 million AI builders.
  • 03Nvidia's open physical AI dataset has been downloaded more than 15 million times and includes 350,000 real and simulated trajectories.
  • 04The dataset also contains 57 million grasps to seed robot training workflows.
  • 05Jetson Thor is being wired into LeRobot's Reachy 2 humanoid to run vision-language-action models on device.

Thomas Wolf, cofounder and chief science officer at Hugging Face, framed the release as the moment robotics gets its shared research infrastructure. He argued that developers now have shared models, data and workflows to train and evaluate robots in the open — the same pattern that accelerated language modeling once Llama and its descendants opened up.

The dataset picture is where the numbers get concrete. Nvidia's open physical AI dataset, which it says is the largest of its kind, has been downloaded more than 15 million times. It contains over 350,000 real and simulated trajectories and 57 million grasps, the kind of volume that a single robotics startup cannot realistically generate on its own. Feeding that into LeRobot's training loop lowers the cold-start cost for any team trying to bring up a new manipulation or humanoid policy.

Simulation is the other half of the pipeline. Isaac Sim and Isaac Lab are wired in for environment setup, data generation and policy validation before code touches a real robot. Isaac Lab-Arena connects into LeRobot's Environment Hub so developers can prototype complex scenes, register them and use them to train and evaluate generalist policies including GR00T, Pi and SmolVLA. The point is that a policy trained by one team can be evaluated by another on the same environment without rebuilding the harness.

And with NVIDIA Cosmos 3 planned next, the community will have a path to bring frontier world models into that same collaborative loop.
Thomas Wolf, Cofounder and Chief Science Officer, Hugging Face

On the deployment side, Nvidia's Jetson Thor edge compute module is being integrated with LeRobot's Reachy 2, an open source humanoid platform. That closes the loop: a developer can now, in principle, collect teleoperation data with Isaac Teleop, fine-tune GR00T 1.7 in LeRobot, validate it in Isaac Lab-Arena, and deploy it to a Reachy 2 running Jetson Thor without leaving open source tooling.

Cosmos 3, the world foundation model Nvidia has flagged for a later LeRobot integration, is meant to generate and augment robotics data and simulate scenarios where real-world capture is too expensive or too slow. Nvidia has been pushing world models hard through the AI Nations strategy this site covered last week, and the LeRobot integration is a consumer-facing extension of the same bet: that synthetic data and simulated rollouts will do for robotics what web-scale text did for language models.

Related · from this week
Nvidia opens Alpamayo 2 Super for commercial robotaxi deployment
Jaeden Schafer · 5 min read →

The open-source framing is doing real work here, but it's not without caveats. GR00T 1.7 is open and commercially viable, but the full stack still assumes Nvidia silicon at both the training and edge-inference layers. A developer running LeRobot on non-Nvidia compute gets access to the datasets and the model weights, but the simulation and deployment tooling is tightly optimized for Isaac Sim, Isaac Lab and Jetson. That is a defensible strategy for Nvidia, and it is also the reason competing frameworks will keep existing.

For the AI market, the interesting shift is that Nvidia is treating robotics the way it treated the LLM stack — sell the compute, seed the ecosystem with an open model, let a partner own the community layer. Hugging Face gets a defensible position as the default hub for open robotics work. Robotics startups get a viable alternative to building everything in-house. And Nvidia collects the tax on all of it at the chip layer, from the Isaac Sim training cluster down to the Jetson Thor on the robot. If the pattern holds, LeRobot becomes the place where open humanoid policies get published, benchmarked and improved — and the humanoid category finally gets the shared substrate that has been missing.

ShareXLinkedInEmail
AI Box

Every AI model. One chat.

The latest models from ChatGPT, Claude, Gemini, Sora, ElevenLabs — 80+ models in a single chat. Compare answers side by side. Pick the best one every time.

  • ChatGPT, Claude, Gemini, Grok, DeepSeek — in one chat
  • Generate images & video with Sora, Veo, Ideogram
  • Compare any two models side by side
  • From $8.99/mo · 80+ models, all included
Try AI Boxaibox.ai
Trusted by 3,000+ teams
Got a tip?

Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.

Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.

AI Box Daily briefingFree · Daily · No fluff

Stay ahead of everyone in AI.

The tightly edited AI news email engineers, founders, and investors actually open. One email. Every weekday. Five minutes to finish.

Loved by 10,000+ AI professionals
Free forever. Unsubscribe with one click.

The briefing read inside teams at

Keep reading

More from Models

Nvidia logo
Models

Nvidia opens Alpamayo 2 Super for commercial robotaxi deployment

The 30B-parameter driving model tops LingoQA against Qwen, Gemini and GPT-4o, and ships under a permissive Linux Foundation license.

Jaeden Schafer5 min read
Nvidia logo
Models

Nvidia unveils Jetson Thor T3000 and T2000 for mainstream robotics

The Blackwell-based T3000 delivers 865 FP4 teraflops in half the size of the T5000; both modules ship Q1 2027.

Jaeden Schafer5 min read
Nvidia logo
Models

Nvidia unveils GraspGen-X, LCDrive and NitroGen foundation models at CVPR

Three new physical AI models target zero-shot robot grasping, faster autonomous-vehicle reasoning, and gameplay-trained embodied agents.

Jaeden Schafer5 min read