Recently, the Shanghai-based robotics startup ZhiYuan Robotics, in collaboration with the Shanghai Artificial Intelligence Laboratory, the National Local Joint Human-Robot Innovation Center, and Shanghai KupaSi, officially launched the open-source million-scale real-world dataset, AgiBot World, aimed at supporting generalized and universal robotic model training. It is reported that this is the world's first million-scale dataset based on comprehensive real-world scenarios, an all-encompassing hardware platform, and full-quality control.
The AgiBot World dataset was born from ZhiYuan's self-built large-scale data collection factory and application experimental base, covering a total area of over 4,000 square meters and including more than 3,000 real objects. It replicates five core scenarios: home, dining, industry, supermarkets, and office, and features over 80 diverse skill videos from daily life. Compared to Google’s Open X-Embodiment, AgiBot World boasts a data scale ten times larger, a scene coverage area expanded by 100 times, and data quality elevated from laboratory-level to industrial-level standards.
A representative from ZhiYuan Robotics stated that the release of AgiBot World will significantly advance humanoid robot technology, enabling robots to engage in various aspects of human daily life rather than just performing simple desktop tasks. It is known that AgiBot World is the third open-source project released by ZhiYuan Robotics this year, with related data to be uploaded in batches on HuggingFace, GitHub, and the AgiBot World project homepage.
In the future, ZhiYuan Robotics plans to gradually open source tens of millions of simulation data to support more generalized and universal model training; release a foundational model for embodiment, which can support model fine-tuning; and provide a complete toolchain to achieve a closed loop of data collection, training, and evaluation.
GitHub link: https://github.com/OpenDriveLab/agibot-world
Project homepage: https://agibot-world.com/
