Reachy2: Hugging Face debuts open-source household humanoid project

In its efforts to promote research in the humanoid domain, the open-source French robotics company collaborated with Hugging Face to train its robot to handle multiple household chores and ensure safe interactions with humans and dogs.

Pollen Robotics has believed in developing open source, open science, and open data goods ever since it launched Poppy, the first 3D printed open-source humanoid robot, back in 2013.

Open-source robot training

The Reachy2 robot was initially controlled by a human operator who used a virtual reality headset to guide it through various tasks, such as placing cups on a dish rack and safely handing objects like an apple to a person.

According to Cadence, who spoke to VentureBeat, a machine learning algorithm subsequently analyzed 50 short videos, each about 15 seconds long, of these VR teleoperation sessions.

Through this analysis, the algorithm learned to perform the tasks independently, instructing Reachy2 on how to execute them. Each video corresponded to different sensors within the Reachy2 robot.

After training for between 40,000 and 60,000 steps, Reachy2 mastered the skill of rotating to hand an apple and returning to its original position.

The firm has open-sourced the dataset on Hugging Face with the robotics firm, along with the model it trained and employed for the demonstration.

Cadene highlights that the information is available for anyone to utilize. “You can do the same at home on smaller robots,” he said on X.

Aiding the research, Pollen Robotics Reachy2 is ideal for such applications, thanks to its all-new 7-DoF bio-inspired arms (7-pound payload each) and AI-ready circuitry, which enable improved perception, interaction, and manipulation.

AI robotics for everyone

Hugging Face, which maintains one of the largest repositories of open-source AI models and codes, launched its Le Robot project in March 2024.

It aims to create an open-source toolkit designed to democratize AI robotics and inspire a new generation of roboticists. It is now available on GitHub.

“The next step of AI development is its application to our physical world. Thus, we are building a community-driven effort around AI for robotics, and it’s open to everyone,” said Cadence in a post on X.

This toolkit is a comprehensive platform offering a versatile library for sharing, visualizing data, and training state-of-the-art models. Users can access a plethora of pre-trained models to jumpstart their projects.

LeRobot also integrates seamlessly with physics simulators, allowing enthusiasts and developers without physical robotics hardware to simulate and test their AI models in a virtual environment.

Examples of the robotic capabilities are available in the LeRobot library on GitHub, including training robots to navigate unmapped spaces and grasp objects from video.

By making it open-source, Hugging Face hopes a global community of developers, researchers, and hobbyists would contribute to and benefit from the collective progress in AI robotics.

“We are building a diverse community from various backgrounds, software and hardware, to develop the next generation of smart robots in the real-world,” said Cadence.