Recently, ZhiYuan Robotics successfully passed the CR-3-06:2025 “Humanoid Robot Dataset CR Product Certification Implementation Rules”, obtaining the industry’s first humanoid robot dataset product CR certification certificate, numbered “001”, becoming one of the first companies in the country to be certified based on the T/SAIAS 024—2025, T/SAIAS 025—2025, T/SAIAS 026—2025, and T/SAIAS 027—2025 series of group standards.
According to reports, this certification system comprehensively evaluated the four core elements of humanoid robot dataset construction: classification and coding specifications, data annotation standards, unified dataset format requirements, and comprehensive quality evaluation criteria.
Looking at the AgiBot World dataset launched by ZhiYuan Robotics, it is the world’s first real machine dataset based on global real scenarios, an all-capable hardware platform, and full quality control, establishing a “data benchmark library” in the field of embodied intelligence.
This dataset achieves a dual breakthrough in scale and quality: the long-range data scale is ten times higher than that of Google’s Open X-Embodiment dataset, covering basic operations such as grabbing and pushing, as well as complex long-range tasks like stirring, folding, ironing, and even memory stick installation and cutlery sorting, with 80% of task durations concentrated between 60-150 seconds, covering multiple sets of atomic skill combinations from basic operations to complex long-range tasks; the scene coverage has expanded 100 times, and the data quality has jumped from laboratory level to industrial level standards.
In terms of scene and task coverage, the AgiBot World dataset has achieved full-dimensional expansion. Relying on ZhiYuan’s self-built data collection factory and application experimental base of over 3000 square meters, the dataset replicates five core scenes: home, dining, industrial, supermarket, and office, including hundreds of real sub-scenes and over 3000 real items.
For example, the home scene replicates the real layout of human residences, including core spaces such as bedrooms, living rooms, kitchens, and bathrooms, enabling robots to perform household cleaning, item organization, and kitchen tasks.
ZhiYuan Robotics states that the high quality of the AgiBot World dataset is attributed to robust hardware support and a rigorous quality control system. The data collection robot, Spirit G1, is equipped with eight cameras arranged in a surround layout, allowing the robot to perceive dynamic changes in the surrounding environment in real-time from all angles.
At the same time, the entire data collection process follows a strict procedure of “multi-party planning – dual screening at the edge and cloud – frame-by-frame manual review – algorithmic secondary verification”, ensuring the standardization and usability of the data from the source.
Industry insiders believe that data has become one of the most significant shortcomings in the current development of the AI industry, especially in fields like embodied intelligence and robotics that deeply interact with the real world, where the lack of high-quality data severely restricts technological iteration and commercialization.
Unlike the data required for large language models, embodied intelligence must deal with highly complex and heterogeneous scenarios in real environments, involving variables such as light, sound, and temperature, and such data is difficult to obtain on a large scale through traditional methods.
Yang Lin, chief analyst of the computer industry at Guotai Junan Securities, previously told the Shanghai Securities News that companies focusing on data production, processing, and trading will face significant opportunities in the future. Currently, local governments, including Shanghai, have begun to provide targeted support for data-related enterprises, while many regions have established big data exchanges to promote the standardized circulation and sharing of high-quality AI data, making this area worthy of market attention.
Author:Sun Xiaocheng
