Replicating the open source approach that has made it successful in corporate AI models, the French flagship is developing a whole range of robots, from conversational robots to robotic arms.
Who better VRP than Jensen Huang? The founder of Nvidia took a liking to Reachy Mini, a cute robot, as tall as three apples (28 cm), immediately attracting sympathy with its asymmetrical eyes and telescopic antennas. At the last edition of CES in Las Vegas, the boss, dressed all in leather as usual, showed how this little R2D2 can become an ideal office colleague with whom to discuss any subject in natural language (see the video).
A real spotlight for Hugging Face’s latest robot. Founded ten years ago by three French people, the company, nicknamed the GitHub of AI, is above all known for facilitating the adoption of open source AI models by companies. Hugging Face has replicated this business model in the world of robots.
In May 2024, the scale-up launched, under the direction of Rémi Cadene, a former engineer of Tesla’s Optimus program, the LeRobot library. Based on PyTorch, this open-source framework is designed to democratize access to robotics learning models and training data. The platform has 21,100 stars on GitHub and is the subject of 3,500 forks.
In April 2025, Hugging Face entered hardware by purchasing Pollen Robotics, a Bordeaux startup of around twenty people created by two former Inria researchers, Matthieu Lapeyre and Pierre Rouanet. With this acquisition, Hugging Face becomes the owner of Reachy 2. Big brother of Reachy Mini, this sophisticated humanoid robot already has the same friendly face but also has a mobile base equipped with three omni-wheels, a LiDAR system and articulated arms allowing it to support loads of up to 3 kg.
Considering its price, around $70,000, Reachy 2 is intended for R&D teams and research laboratories. Another advantage is that it is teleoperable. Using a virtual reality headset, an operator can control the robot’s actions and even teach it new movements. “This makes it possible to train it to make it more autonomous according to contextual analysis,” explains Rémi Fabre, robotics engineer at Pollen Robotics.
Offered in two configurations – a lite version by connecting it to a USB-C cable (around $300) and a wireless version with onboard Raspberry Pi 5 (around $500), Reachy Mini goes much further in the democratization of robotics by targeting the maker population. “It takes two hours to assemble it piece by piece,” continues Rémi Fabre. “The user not only feels a sense of pride in having built it themselves but they also know, in complete transparency, what elements make up the robot. A passionate parent will, for example, assemble the robot with their child.”
Endow a robot with emotions and cognitive abilities
Pollen Robotics also cultivates a sense of empathy at Reachy Mini. “When the robot speaks, it looks at its interlocutor and oscillates its head in three directions,” explains Rémi Fabre. “This emotional charge makes the interactions more lively.” We then speak of cognitive robotics, or the ability to play this or that emotion.
But what to do with this conversational robot? Reachy Mini is associated with an application store, which can be downloaded directly from its dashboard. He thus transforms himself into a chess teacher or an English coach. It can also play “one, two, three suns” or become a home automation hub to control the temperature or the roller shutters in your home. “A user can create an application without software skills, such as, for example, transforming the robot into a game controller, then sharing it,” adds Rémi Fabre.
Hugging Face’s family of robots isn’t limited to the two Reachy brothers. Still in its desire to make robotics accessible, the company has in its catalog, SO-101, an open source robotic arm, sold between 100 and 500 dollars depending on the assembly level, specifically intended for AI developers. Although it of course does not offer the same power as an industrial robot, it is possible, thanks to reinforcement learning, to teach it to accomplish specific tasks.
Embodying open source physical AI
These open source machines allow us to experience the contribution of generative AI applied to robotics. What we still call physical AI, is the ability for a machine to “perceive” the real world and interact with it. Derived from LLMs, new models provide robots with real cognitive capabilities. By learning and not by programming, they execute complex orders and adapt to their environment.
With SmolVLA, Hugging Face has developed its own Vision Language Action (VLA) model, trained on community data from its LeRobot platform. A VLA model translates visual and textual data into motor commands that allow the machine to perform a series of actions. Unlike an LLM calling cloud services via APIs, this type of model operates locally, which requires ad hoc onboard resources.
“It is an ideological opposition between the use of a generic foundation model or a set of specific models,” comments Rémi Fabre. “Inference in local mode avoids sending data to the cloud. This reduces dependence on digital giants and allows us to maintain control of the data.”
By wishing to embody open source physical AI, Hugging Face can, in fact, offer an alternative to the proprietary models offered by the Americans Tesla Optimus and Boston Dynamics or the Chinese Xiaomi and Unitree. A significant sovereignty issue in the current geopolitical context.
To meet this challenge, Hugging Face has created an ecosystem. A year ago, the company entered into a strategic collaboration with Nvidia to combine the LeRobot framework with the American giant’s robotics technologies, Omniverse and Isaac Lab. These offer GPU-accelerated synthetic data generation and simulation capabilities. The objective being to be able to train and test robotic models before deploying them on the machine.
Launch of UMA in the wake of Hugging Face
Finally, robotics activity goes beyond the scope of Hugging Face alone. Last December, two co-founders of LeRobot, Rémi Cadène and Simon Alibert, alongside Pierre Sermanet (ex-Google DeepMind) and Robert Knight (ex-The Robot Studio) created the company UMA, for Universal Mechanical Assistant, with the support of AI luminaries like Yann LeCun (AMI) and Thomas Wolf, chief science at Hugging Face and a certain Xavier Niel for business angel.
Based in Paris, UMA aims to design advanced AI in the physical world and build humanoid robots ready to work in real environments, at scale. The young company intends to meet the challenges of robotics in the fields of logistics, health and home care for the elderly.
UMA plans to develop two complementary robot models: a mobile industrial robot equipped with two arms for warehouses or assembly lines and a humanoid robot capable of moving among humans and collaborating with them. Quite a program.
