AgiBot World - A million-device dataset open-sourced by Zhiyuan Robotics

What is AgiBot World?

AgiBot World is an open-source dataset of millions of real-machine data from Logic Robotics, aiming to advance the development of embodied intelligence. The dataset includes over eighty everyday skills, covering five core scenarios: home, dining, and industry. Its scale and quality far surpass Google's Open X-Embodiment. Based on Logic Robotics' self-built data acquisition factory and experimental base, and utilizing advanced hardware configurations such as eight cameras and a six-DOF dexterous hand, high-quality data acquisition across all real-world scenarios has been achieved.

Main functions of AgiBot World

• Diverse task coverageThe AgiBot World dataset includes more than 80 diverse skills used in daily life, ranging from basic operations such as grasping, placing, pushing, and pulling, to more complex actions such as stirring, folding, and ironing, covering almost all the action requirements of daily life.
• Full-domain realistic sceneThe dataset was created in the large-scale data collection factory and application experimental base built by Zhiyuan Robotics. The total space covers more than 4,000 square meters and contains more than 3,000 real objects, replicating five core scenarios: home, catering, industry, supermarket and office, providing robots with a highly realistic production and living environment.
• All-round hardware platformThe robotic platform used for dataset collection is equipped with an eight-camera surround layout, enabling real-time, all-around perception of dynamic changes in the surrounding environment. The robot also possesses a dexterous hand with six degrees of freedom, capable of performing various complex operations such as ironing. With a maximum of 32 degrees of freedom throughout its body, and equipped with a six-dimensional force sensor and high-precision visual-tactile sensors at its end effector, the robot can perform delicate tasks with remarkable efficiency.
• Quality control throughout the entire processDuring the data collection process at AgiBot World, Zhiyuan Robotics adopted a multi-level quality control and on-the-loop human intervention strategy. From the professional training of data collectors to the strict management of the data collection process, and then to the screening, review and labeling of data, every step was carefully designed and strictly controlled.
• Dataset contentAgiBot World includes over eighty diverse skills used in daily life, ranging from basic operations such as grasping, placing, pushing, and pulling, to complex actions such as stirring, folding, and ironing, covering almost all the action needs of human daily life.
• Open source projectThe AIZE Robotics plans to gradually open-source tens of millions of simulation data sets to support the training of more generalized and universal large models; it will release a large model with an embodied base that can support model fine-tuning; and it will release a complete toolchain to achieve a closed loop of data collection, training, and evaluation.

AgiBot World project address

• Project official website:agibot-world.com
• Github repository:https://github.com/OpenDriveLab/agibot-world
• HuggingFace model library:https://huggingface.co/agibot-world

Application Scenarios of AgiBot World

• Home SceneAgiBot World recreates the layout of a real human home, including core spaces such as bedrooms, living rooms, kitchens, and bathrooms. In these scenarios, robots can perform household chores such as cleaning, organizing items, and kitchen tasks.
• Dining sceneThe dataset includes food-related tasks, such as stirring and folding in the kitchen, as well as possible restaurant service tasks.
• Industrial SceneThis system simulates sorting and logistics automation, replicating industrial warehouses and production lines, including sorting systems, packaging equipment, and conveyor belts. This helps train robots to perform tasks such as material sorting, packaging, and logistics handling.
• Supermarket sceneIt highly replicates the layout of supermarket shelves and checkout areas, including sections for fresh produce, daily necessities, frozen goods, and more. This helps train robots to simulate tasks such as shelving, inventory management, customer guidance, and unmanned checkout.
• Office SceneIt covers tasks that robots might perform in an office environment, such as document organization and item delivery.