At the Machina show in Paris, Nvidia and Hugging Face announced the integration of several Nvidia technologies into LeRobot, Hugging Face’s open source platform dedicated to robotics.
The objective of this partnership between Hugging Face and Nvidia is to accelerate the development of open source robotics by giving developers access to tools covering the entire development cycle, from data collection to deployment. Among them are Isaac Teleop, a teleoperation and data collection tool for real or simulated robots, as well as GR00T N1.7, NVIDIA’s latest foundation model for robotics, now available under a license compatible with commercial use. We met Caroline Pascal, robotics engineer at Hugging Face, and Sasa Docca, Product Marketing Robotics manager at Nvidia, to discuss the ambitions of this collaboration and their vision of the future of open source robotics.
JDN. What does the partnership you have just announced consist of and what does it concretely bring to developers?
Sasa Docca (Nvidia). Our collaboration with Hugging Face aims to give developers everything they need, from data collection to deployment. Hugging Face today brings together one of the largest open source communities in robotics.
Concretely, our announcement is based on two pillars. The first is the integration of Isaac Teleop, a teleoperation and data collection tool for both real and simulated robots. The second is GR00T N1.7, the latest version of our foundation model for robotics, now available under a license compatible with commercial use.
We are seeing more and more companies looking to move from proof of concept to industrial deployment. With Isaac Teleop, GR00T N1.7 and the datasets published on Hugging Face, they now have a complete ecosystem to develop and deploy their robotic applications.
Caroline Pascal (Hugging Face). As Thomas Wolf, co-founder of Hugging Face, often reminds us, open source consists of making the most advanced research accessible so that everyone can learn and develop their own projects. Open source accelerates innovation: developers experiment more quickly, share their results and the entire community benefits.
Who exactly is this solution aimed at?
Caroline Pascal. It is aimed at students and researchers as well as large companies, startups and independent developers. Beyond industrial robotics, it may also be of interest to players who develop service or cleaning robots. GR00T and Isaac Teleop offer them a technological base to quickly experiment and create their own robotic applications.
Sasa Docca. Developers are at the heart of this ecosystem. Our foundation model has already been pre-trained on huge volumes of data from the Internet, human demonstrations and synthetic data. All that remains is to adapt it to each use case using fine-tuning. We take care of the most expensive part of training so everyone can focus on their applications.
GR00T is not reserved for humanoid robots: it works on a wide variety of robotic platforms, from industrial robots to biomedical robots.
Finally, Isaac Teleop allows data to be collected and shared on Hugging Face. By opening up both models and datasets, we promote more collaborative development of robotics.
Does data remain the main challenge for building robots today?
Sasa Docca. Yes. Our foundation model was trained on over 20,000 hours of egocentric data, supplemented with data from simulations and real robots. No developer could bring together such a volume alone. The advantage of open source is that the model, but also the training data, are accessible. This provides greater transparency, control and security. Data collection can come from teleoperation, human demonstrations or simulations. All these sources are essential to train models capable of evolving in the real world.
This is precisely the role of Cosmos 3, our world foundation model. It combines real data and simulation to generate new environments, enrich datasets or even evaluate robotic policies. We are currently working with Hugging Face on its integration. Ultimately, a large part of the data necessary for training robots will be automatically generated by these models. But this is only possible if they remain firmly grounded in the laws of physics. Without this anchor, they begin to hallucinate.
Caroline Pascal. This is precisely what excites me about Cosmos 3. It can generate new scenarios while remaining consistent with the laws of physics. No matter how much real data you collect, there will always be blind spots. “World models” make it possible to fill these gaps and make robots safer by teaching them to better anticipate the situations they will encounter. For me, the integration of Cosmos 3 will be a major milestone for all robotics.
Why is open source so important in robotics in your eyes?
Sasa Docca. When a technology is open, we can understand how it works, adapt it and trust it more. In robotics, value is created mainly during post-training: everyone can adapt a foundation model to their own use case, whether it is a simple pick-and-place task or a much more complex manipulation. Open models then make it possible to share these improvements and enrich them collectively. This is how innovation progresses. If the models are not open, we simply cannot innovate at the same pace.
Caroline Pascal. There is also a security issue. Would you trust a robot operating with a completely opaque model, without knowing what’s going on inside?
With open source, everything is visible. If a problem exists in the model or in the code, eventually someone will catch it. This transparency is essential to ensure security, privacy and trust.
Why did Hugging Face choose to focus on robotics?
Caroline Pascal. We cannot do AI only on a screen. After text, image and video, the next step is naturally embodied AI, i.e. robotics. This is also an area where open source is particularly important. If proprietary systems are deployed in the real world tomorrow, we must be able to understand how they work and ensure that they are secure. That’s why it’s essential to lay the right foundations today.
Why is NVIDIA investing so much in physical AI?
Sasa Docca. Physical AI represents a huge opportunity. Robotics is not limited to humanoids: an autonomous vehicle, a robotic arm, a smart camera or a building capable of self-regulating are already robotic systems. Our ambition is to provide a complete platform, from hardware to AI models, while giving developers the freedom to use only the building blocks they need. Security is another major issue. With NVIDIA Halos for Robotics, we are adapting safety technologies developed for autonomous vehicles to robots to support them from training to deployment.
Do you think that one day we will achieve a form of AGI applied to robotics?
Sasa Docca. We are already on this trajectory. We are gradually moving from specialized robots to general robots, then tomorrow to “specialist generalists”, capable of quickly learning new tasks. Like humans, robots progress through experience. The more data they accumulate from the real world, the smarter they become. Pre-training is essential, but it’s deployment that provides the most valuable experiences.
Caroline Pascal. Models must also learn from their mistakes. Robotics introduces new challenges, with highly multimodal data and actions to be generated rather than simple texts or images. Openly sharing results allows the entire community to more quickly identify the most promising approaches and accelerate progress.
Sasa Docca. We are also seeing the convergence between agentic AI and physical AI. Agents will be able to automatically generate thousands of simulation or data collection scenarios that a human would not have thought of. This combination of AI agents, simulation and world models will accelerate the development of the next generation of robotic intelligence.
How do you judge the European robotics ecosystem today?
Caroline Pascal. He is booming. Europe benefits from a solid ecosystem, where startups collaborate easily while being stimulated by competition with the United States or China. Europe already has the foundations of a very strong robotics ecosystem.
Sasa Docca. At NVIDIA, we work with partners all over the world, but Europe looms large thanks to its long industrial tradition. We are seeing a new generation of very promising companies emerge, such as 1X, NEURA Robotics or Schaeffler, all of which are partners with NVIDIA. The partnership with Hugging Face is fully in line with this dynamic.
