Humanoid robots are trained to swing their arms at the Zhejiang humanoid robot innovation center in Ningbo, east China's Zhejiang Province, Feb. 19, 2025. (Photo: Xinhua)
Shanghai-based robotics firm Fourier Intelligence has released its open-source full-size humanoid robot dataset, Fourier ActionNet, along with the world's first full-process tool-chain.
The move aims to enhance AI robot training and offer a comprehensive solution for developers and research institutions worldwide, fostering innovation and collaboration among global robotics community, according to an announcement seen on the company's WeChat account on Monday.
The initial dataset includes over 30,000 high-quality real-machine training entries, covering dexterous hand movements and specialized imitation learning data for hand-related tasks, covering application scenarios such as picking up and laying down tools, performing household tasks as well as other actions.
High-quality robot motion data is essential for advancing embodied intelligence, said a robotics expert, adding that open-sourcing data allows researchers and developers to share insights and methodologies, fostering a better understanding of how embodied systems learn and adapt. This collaboration improves robot adaptability in dynamic environments and strengthens human-robot interaction.
However, collecting real-world robot motion data has been constrained by high costs and low annotation accuracy, limiting industry progress, Jiang Lei, chief scientist of Shanghai-based National and Local Co-Built Humanoid Robotics Innovation Center, told the Global Times on Monday.
Jiang noted that fostering an open-source culture in embodied intelligence can spur innovation, drive academic breakthroughs, and make transformative products by leveraging global research expertise.
Currently, Fourier has collaborated with over 20 top domestic and international research institutions and leading industry enterprises. Based on its self-developed humanoid robot platform, they have achieved several breakthroughs in research areas such as reinforcement learning, imitation learning, Visual-Language-Motor large models, and perception systems. In the future, Fourier will continue to release more advanced data modules covering full-body motion control and multi-task coordination, said the company.
A group of Chinese robotics startups have unveiled a vast dataset designed to fast-track the widespread deployment of autonomous humanoid robots. Among them, Shanghai-based AgiBot has launched its dataset, named AgiBot Digital World, which comprises more than 1 million robot action trajectories captured from 100 robots in real-world scenarios. This dataset originates from a large-scale data collection facility and experimental base established by AgiBot, also known as Zhiyuan Robotics, according to local media outlet Shanghai Observer.
Hangzhou-based robotics firm Unitree has confirmed to the Global Times that it has updated its H1 and G1 humanoid robots with an open-source full-body dataset. Unitree said that they used motion capture data to make their movements look more natural and smoother.
"By offering a comprehensive, open-source dataset spanning multiple industries and scenarios, companies are making advanced robotic learning more accessible." This means that, instead of expensive, exclusive research being limited to a few well-funded labs, now more people can contribute to and learn from cutting-edge robotic data, said Jiang.
