NVIDIA and Hugging Face join forces in LeRobot: Isaac GR00T 1.7, Isaac Teleop, and Cosmos 3 arrive in open source for robotics

NVIDIA and Hugging Face join forces in LeRobot: Isaac GR00T 1.7, Isaac Teleop, and Cosmos 3 go open source for robotics

🔎 The ChatGPT moment of robotics just happened in open source

Humanoid robotics is going through a tipping point. Not in five years, not in a closed lab — now, on Hugging Face, with a model downloadable by any developer.

On July 7, 2026, NVIDIA and Hugging Face announced the integration of Isaac GR00T N1.7, Isaac Teleop, and the upcoming Cosmos 3 into the LeRobot ecosystem. This is a paradigm shift: for the first time, a foundation model for humanoid robots is open and commercially licensed.

Behind the technical announcement lies a staggering number. NVIDIA is connecting its 3 million robotics developers to the 16 million active builders on Hugging Face. The gap between AI research and physical robotics has just been reduced to a git clone.

This integration is part of a broader movement that NVIDIA partly triggered with NVIDIA Cosmos 3 et Isaac GR00T : le ChatGPT moment de la robotique. Except that today, it is no longer a keynote promise — it is code, model weights, and publicly available datasets.

The Essentials

• Isaac GR00T N1.7: first open-source and commercially viable VLA (Vision-Language-Action) model for humanoid robots, 3 billion parameters, available on Hugging Face.
• Isaac Teleop: standardized human demonstration collection tool, with interoperable formats directly shareable in LeRobot.
• Cosmos 3: world foundation model planned for robotics data generation and augmentation, integrated into the LeRobot pipeline.
• Massive dataset: 350,000 real and simulated trajectories, 57 million grasps, over 15 million downloads.
• Open-source end-to-end pipeline: data collection → training → evaluation → deployment, all within a unified ecosystem.
• Hardware deployment: integration with Jetson Thor and Reachy 2 for open-source humanoid robots.

Recommended Tools

What Isaac GR00T N1.7 really is — and why it's unprecedented

Isaac GR00T N1.7 is a 3-billion-parameter VLA (Vision-Language-Action) model designed for general-purpose humanoid robots. A VLA model takes an image (what the robot sees) and a natural language instruction (what it needs to do) as input, then directly generates motor actions.

What makes it unique is its dual nature: open and commercially viable.

Before GR00T N1.7, the situation was at a standstill. Foundation robotics models existed among tech giants — Google with RT-X, Tesla with Optimus — but remained closed or inaccessible. Open models like early academic work lacked licenses allowing commercial use. Startups had to rebuild everything from scratch.

GR00T N1.7 breaks this dynamic. As confirmed by the model page on Hugging Face, the model is released under an open commercial license. A developer can download it, post-train it for their specific robot, and deploy the result in a commercial product. It's a first in the field of foundation robotics models, as discussed on the NVIDIA developer forums.

The architecture is cross-embodiment with integrated reasoning. In practice: the same model can control robots of different morphologies — a robotic arm, a bipedal robot, a wheeled torso — because it reasons about the task rather than mechanically reproducing movements.

The source code is available on GitHub (NVIDIA/Isaac-GR00T), with complete documentation and deployment examples.

Isaac Teleop : standardizing robotic data collection

A VLA model is worthless without quality data. This is the current number one bottleneck in robotics. Isaac Teleop addresses precisely this problem.

Isaac Teleop is a tool that allows a human to teleoperate a robot to collect demonstrations. The key: data is collected in a standardized, interoperable format, and is directly shareable within the LeRobot ecosystem.

Before Isaac Teleop, every robotics lab had its own data format. Datasets were incompatible with each other. A team could not combine its data with that of another lab without colossal conversion work.

With the integration into LeRobot described in the dedicated Hugging Face blog post, the demonstrations collected via Isaac Teleop follow a unified schema. They directly feed into the GR00T N1.7 training workflows.

The process has become almost trivial: a human operator teleoperates the robot → trajectories are recorded in the LeRobot format → they are pushed to the Hugging Face Hub → anyone can use them to fine-tune GR00T N1.7 on a new task or a new robot.

It is the equivalent of the moment when ImageNet standardized computer vision datasets in 2009. Except here, we are not talking about static images, but complex sensorimotor sequences.

The open dataset: 350K trajectories, 57M grasps, 15M downloads

The figures for the dataset integrated into LeRobot are spectacular. The NVIDIA announcement cites 350,000 real and simulated trajectories, 57 million grasps, and over 15 million downloads.

To put these numbers into perspective: Google's RT-1 dataset, considered a milestone in 2022, contained 130,000 episodes. LeRobot far exceeds this scale, along with added diversity in modalities (real + simulated) and robot morphologies.

The 57 million grasps are particularly significant. Object manipulation — grasping, turning, inserting, assembling — remains one of the most difficult challenges in robotics. Having this many grasp examples in an open and standardized dataset changes the game for training grasping policies.

The 15 million downloads mainly indicate one thing: the community is hungry for robotics data. LeRobot has already become the go-to hub, and NVIDIA's contribution will accelerate this momentum exponentially.

This dataset is not a raw dump. It is organized, documented, and designed to be directly usable in GR00T N1.7 training pipelines via LeRobot. It is a data infrastructure, not just a file storage.

The end-to-end pipeline: from data to deployment, all open source

The most structural contribution of this integration is not an individual model or tool. It's the complete pipeline.

The LeRobot workflow with NVIDIA components now covers the entire development cycle of an intelligent robotic system:

Collection — Isaac Teleop allows for human demonstration of tasks, with standardized formats directly integrated into the Hub.

Training — GR00T N1.7 serves as the foundation. Developers can post-train it on their specific data via LeRobot workflows, as detailed in the Hugging Face blog on GR00T N1.7. The model adapts to new embodiments and tasks without starting from scratch.

Evaluation — Isaac Lab-Arena, integrated into the LeRobot Environment Hub, provides simulation environments to test learned policies before any physical deployment. This is crucial for safety and iteration speed.

Deployment — The pipeline culminates on real hardware, notably Jetson Thor (NVIDIA's robotics chip) and Reachy 2 (Pollen Robotics' open-source humanoid robot).

This pipeline has existed among major robotics players for years. But it was proprietary, expensive, and inaccessible to 99% of developers. Making it open source means democratizing access to world-class industrial infrastructure.

Cosmos 3: the world foundation model that will multiply data

Cosmos 3 is announced as planned for LeRobot integration, and it is potentially the most disruptive component of the bunch.

A world foundation model (WFM) is a model that understands and generates realistic physical scenes. Cosmos 3, in the context of robotics, will be used to generate and augment synthetic training data.

The problem is simple: collecting real robotic data is slow, expensive, and difficult to scale. A human must physically teleoperate a robot for hours to capture just a few hundred trajectories.

Cosmos 3 promises to solve this problem by generating ultra-realistic simulated scenes where the robot can train virtually. The model understands the laws of physics, interactions between objects, and spatial constraints — which allows it to create relevant training scenarios.

The link with NVIDIA Cosmos 3 et Isaac GR00T is direct: Cosmos 3 is the data generation engine that powers GR00T N1.7. The two form a coherent system where the WFM produces the data and the VLA leverages it.

The promise: going from thousands of real trajectories to millions of realistic ones, without ever touching a physical robot. This is what robotics has been missing to cross the threshold that NLP crossed with LLMs when datasets went from a few million to billions of tokens.

Jetson Thor and Reachy 2: Deployment on Open Source Robots

Software without hardware is just a demonstration. NVIDIA thought about the end of the chain.

The LeRobot integration includes explicit support for Jetson Thor, NVIDIA's chip dedicated to humanoid robotics. Thor is designed for real-time inference on VLA models like GR00T N1.7, with hardware optimizations for critical robotics latencies.

But the strongest signal is the partnership with Reachy 2, the open source humanoid robot developed by Pollen Robotics. Reachy 2 is an accessible (compared to industrial humanoids), modular robot whose community is already active in the open source ecosystem.

By making GR00T N1.7 deployable on Reachy 2 via Jetson Thor, NVIDIA and Hugging Face are creating a complete open source stack: a brain (GR00T N1.7), a nervous system (Jetson Thor), and a body (Reachy 2). Everything is modifiable, studyable, and improvable by the community.

This is a fundamentally different approach from that of Boston Dynamics or Tesla, who keep their stack entirely proprietary. Here, any university lab, any startup, can reproduce the pipeline and adapt it. The parallel with open source AI agents with Ollama locally is striking: the same logic of democratization through open source applies, but this time to the physical world.

3 million + 16 million: the merger of communities changes everything

The most strategic figure in the announcement: 3 million NVIDIA robotics developers connected to 16 million Hugging Face builders.

The traditional robotics community is skilled but relatively small. The AI/Hugging Face community is massive but focused on software — NLP, vision, audio. By merging these two ecosystems via LeRobot, NVIDIA creates a sudden network effect.

A developer who knows how to fine-tune LLMs on Hugging Face can now apply the exact same skills to a robotics VLA model. The tools are the same (Hub, datasets, transformers, evaluation). Only the domain changes.

This leverage effect is considerable. If even 1% of the 16 million Hugging Face builders take an interest in robotics thanks to this integration, that makes 160,000 new robotics developers — more than the entire current population of the field.

This is also a strong signal for the governance of agentic AI. As AI agents move from software to the physical world — from chatbots to robots — safety and governance questions become existential. Discussions around the Agentic AI governance led by Google and SAP take on a concrete dimension when agents control robotic arms in the real world.

The place of GR00T N1.7 in the AI model landscape

GR00T N1.7 is not an LLM. It is a specialized VLA model for robotics. But its position in the broader AI ecosystem deserves to be contextualized.

The dominant LLMs in 2025-2026 — GPT-5.5 (agentic score 98.2), Gemini 3 Pro Deep Think (95.4), Claude Opus 4.7 Adaptive (94.3) — excel in verbal reasoning and logical planning. But they cannot drive a motor or bend a robot finger.

GR00T N1.7 occupies a different niche: sensorimotor reasoning. It doesn't debate philosophy, but it understands that a cup on a table can be grasped from above, that the approach must be lateral if an obstacle is present, and that the gripping force depends on the estimated weight of the object.

The most relevant parallel is with open-source models like NVIDIA Nemotron 3 Ultra 550B, which have shown that open models can rival proprietary models. GR00T N1.7 applies this same logic to the robotics domain.

The difference in size (3B parameters for GR00T N1.7 compared to 550B for Nemotron) reflects the nature of the problem. Robotics requires very high-frequency inferences (often 10-50 Hz) with strict latency constraints. Giant models are unusable in this context. Efficiency per parameter takes precedence over brute force.

Impact on the robotics industry and research

For robotics startups, GR00T N1.7 is a game changer. Until now, a startup had to invest months — or even years — to train a base VLA model before even being able to work on its differentiation. With GR00T N1.7 as a starting point, the barrier to entry collapses.

Post-training becomes the new center of gravity. Instead of training a model from scratch, teams fine-tune GR00T N1.7 on their specific data: their robot, their environment, their target tasks. This is exactly the model that succeeded in NLP with LLMs — and it is finally translating to robotics.

For academic research, access to a commercially viable foundation model opens up technology transfer possibilities that did not exist before. A lab can publish a paper with a model derived from GR00T N1.7, and a startup can pick it up and commercialize it without renegotiating licenses.

For robotics giants (Boston Dynamics, Figure, Tesla), the signal is clear: the proprietary moat on robotics foundation models is disappearing. Value will shift toward hardware, proprietary data, and system integration — not toward the base model.

What's still missing — and the current limitations

Despite the scale of the announcement, let's be precise about what isn't solved yet.

GR00T N1.7 is a foundation model, not a ready-made robot. It requires post-training, calibration on the target hardware, and a non-trivial deployment infrastructure. A team without robotics expertise is not going to deploy a functional humanoid in a weekend.

Cosmos 3 is planned, not available. Large-scale synthetic robotics data generation remains a promise, not a product. World foundation models are making impressive progress, but the physical fidelity of generated scenes is not yet sufficient for training without any real data.

Deployment on Reachy 2 is a proof of concept. Moving from a controlled demo to a reliable robot in unstructured environments (an apartment, a warehouse, a construction site) remains a major challenge that goes beyond the software model alone.

Physical safety is the blind spot of open source in robotics. An LLM that hallucinates produces an absurd sentence. A VLA that hallucinates can break an object, injure someone, or damage the robot itself. GR00T N1.7's open commercial license does not solve the liability issue in the event of an accident.

❌ Common mistakes

Mistake 1: Confusing GR00T N1.7 with an LLM

GR00T N1.7 is a VLA model, not a chatbot. It takes visual and language inputs and produces motor actions as output. Comparing it to GPT-5.5 or Claude Opus 4.7 makes no sense — these are different categories with different constraints (latency, inference frequency, input/output modalities).

Mistake 2: Believing the model works out of the box on any robot

GR00T N1.7 is cross-embodiment, but that doesn't mean it's plug-and-play. It requires post-training adapted to the target robot's specific morphology, its sensors, and its action space. Ignoring this step leads to poor or dangerous results.

Mistake 3: Neglecting the quality of teleoperation data

Isaac Teleop standardizes the format, not the quality. Poorly executed demonstrations — jerky movements, excessive human corrections, poorly defined tasks — will produce a poorly trained model, even with perfect infrastructure. Data collection quality remains demanding human work.

Mistake 4: Waiting for Cosmos 3 to start

Cosmos 3 is a future accelerator, not a prerequisite. The 350,000 trajectories already available in LeRobot are sufficient for many post-training use cases. Waiting for a future tool to start working with GR00T N1.7 is a scheduling mistake.

❓ Frequently Asked Questions

Is GR00T N1.7 really free for commercial use?

Yes. The model is released under an open commercial license on Hugging Face, a first for a foundation robotics model. You can download, modify, and integrate it into a commercial product without paying a license fee to NVIDIA. The exact terms are in the LICENSE file of the GitHub repository.

What hardware power is required to run GR00T N1.7?

The model has 3 billion parameters, making it relatively lightweight. The target inference is Jetson Thor, but post-training requires more powerful GPUs (typically an NVIDIA datacenter-class GPU). Light fine-tuning can be done on a standard workstation GPU.

How is LeRobot different from other robotics frameworks?

LeRobot is an end-to-end pipeline integrated with the Hugging Face Hub. The key difference: datasets, models, and environments are all in the same place, in the same format, and interoperable. No need to convert between five different formats to go from collection to deployment.

Will Cosmos 3 replace real data collection?

Not in the short term. Cosmos 3 is designed to augment real data, not completely replace it. Synthetic data still lacks physical fidelity for some complex scenarios. The real + synthetic combination is the most realistic strategy for the coming years.

Can a web developer get into robotics with these tools?

Partially. LeRobot workflows are familiar to anyone who has used Hugging Face (datasets, models, Hub). But robotics adds physical, hardware, and safety constraints that require specific skills. It's a drastically lower entry point, not an elimination of the learning curve.

✅ Conclusion

The integration of Isaac GR00T N1.7 into LeRobot marks the moment when humanoid robotics joins the rest of AI in the open-source era. The complete pipeline — from data collection with Isaac Teleop to deployment on Jetson Thor and Reachy 2 — is now accessible to anyone who knows how to use Hugging Face. The first open and commercially viable foundation robotics model is no longer a promise: it is on the Hub, and it weighs 3 billion parameters. If you want to understand how foundation models are transforming the physical world, start by exploring the GR00T N1.7 repository on Hugging Face.