Shanghai-based startup AgiBot recently announced the launch of what is claimed to be the largest robotic learning dataset, aimed at advancing the development of foundational AI models and enabling robots to perform human-like activities. The dataset, named "AgiBot World," contains over one million training samples from 100 robots, covering a wide range of practical tasks.
Since its founding in 2023 by Zhihui Peng, the company has been dedicated to developing and innovating humanoid robot technology, competing with companies like Tesla and Boston Dynamics in the creation of robots capable of performing basic household tasks such as laundry and folding clothes. Addressing the challenges of current robotic learning benchmarks, especially those related to poor data quality and limited perception, AgiBot's dataset aims to improve the situation by providing large-scale, open-source data.
The AgiBot World dataset includes tasks from homes, restaurants, industrial environments, offices, and supermarkets, with proportions of 40%, 20%, 20%, 10%, and 10% respectively. These data cover fine motor control, tool usage, and multi-robot collaboration. Compared to other public datasets, such as Google's Open X-Embodiment, AgiBot World offers more extensive long-term navigation data, a greater variety of scenarios, and a stronger focus on real-world training data collected in industrial settings.
To ensure the quality and accuracy of the data, particularly at the task level, AgiBot has established specialized data collection facilities to capture real-world data from various industries, households, and daily tasks. Additionally, the company can provide customized data services for specific enterprises upon request, supporting the training and fine-tuning of robot AI models.
The AgiBot World dataset is now available on HuggingFace and GitHub, accessible to robot AI developers and academic researchers for building their own foundational AI models. AgiBot has already launched humanoid and wheeled robots designed for both industrial and household applications.
