On the morning of November 10, the National Data Standardization Technical Committee opened its second “Standards Week” of 2026 in Xiamen. The event ran for four days. Five announcements were made at the opening ceremony: four established working groups for specific industries, while the fifth focused on embodied AI.

Four Industry Groups Draw Leaders From Their Regulators; WG7 Gets an Academician

The National Data Standardization Technical Committee, or SAC/TC609, is overseen by the National Data Administration. It added five groups in one batch, expanding its standards-drafting working groups from six to seven and its industry application groups from six to 10. Four cover meteorology, healthcare, electric power and commerce. The fifth, designated WG7, is the Embodied AI Data Standards Drafting Working Group.

Viewed together, the five lists of group leaders reveal one notable mismatch.

The meteorology group is led by Xiao Wenming, director of the National Meteorological Information Center. The healthcare group is headed by Zhao Wei, director of the National Health Commission’s Center for Health Statistics and Information. The electric power group is led by Fan Pengzhan, deputy director of the State Grid Digitalization Department. The commerce group is headed by Xu Xingfeng, Party secretary of the China International Electronic Commerce Center. Their titles and sectors differ, but all four are officials responsible for data within their respective industry oversight systems.

The embodied AI group is led by Zhou Zhihua, an academician of the Chinese Academy of Sciences and vice president of Nanjing University. Its deputy head is Fan Kefeng, vice president of the China Electronics Standardization Institute.

Each of the four industries has a clearly designated government regulator that can appoint a representative. Embodied AI does not. That is not an oversight but a reflection of the field’s current position: the Ministry of Industry and Information Technology oversees the robotics industry, while the National Data Administration handles data aggregation, rights confirmation, circulation and development as a factor of production. Embodied AI sits between the two. Each agency governs part of the chain, but neither controls it end to end. Appointing an academician specializing in foundational machine learning research uses scholarly authority to anchor the group.

National Data Administration Director Liu Liehong outlined the group’s remit at the opening ceremony: begin preliminary standards research on data collected from physical robots, human demonstration data and context protocols. Of the three areas, two concern data and one concerns protocols.

Three Standards Remain in Draft After 200 Days

Turn the calendar back.

The National Public Service Platform for Standards Information already lists three embodied AI data standards under TC609, with consecutive registration numbers: 20262582-Z-907, “High-Quality Datasets: Embodied AI Data Sources and Constituent Elements”; 20262583-Z-907, “High-Quality Datasets: Specifications for Generating and Processing Simulation and Synthetic Data for Embodied AI”; and 20262584-Z-907, “High-Quality Datasets: Specifications for Data Collection and Model Training at Embodied AI Training Facilities.”

All three were registered on April 24, 2026, with 12-month development cycles, and list the National Data Administration as the responsible authority. When the working group was formally established on November 10, all three were still marked “being drafted.” Two hundred days had passed between registration and the group’s launch.

Based on the 12-month schedule, the three projects are due around April 2027. The working group’s formation at this point suggests it will likely take on projects of this kind: standards that have already been initiated but need an organizational vehicle to move forward. Officials have not confirmed any direct link between the group and these projects; the two dates are simply presented here side by side.

The Two August Standards Fall Outside the Data Administration’s Remit

Work on the same subject had started earlier under a different committee.

Under the National Information Technology Standardization Technical Committee, or SAC/TC28, GB/Z 218.1-2026, “Artificial Intelligence—Embodied AI Data Quality Specifications—Part 1: Real-World Data,” and GB/Z 220-2026, “Artificial Intelligence—Technical Requirements for Embodied AI Data Generation Platforms,” were approved and published by the State Administration for Market Regulation and the Standardization Administration of China on August 27, 2026. Listed as items 11 and 12 in National Standards Announcement No. 37 of 2026, they took effect upon publication. They are China’s first two national standardization guiding technical documents for embodied AI. Such documents are neither mandatory nor recommended standards; their significance lies in establishing common terminology for the industry.

Comparing the two drafting rosters makes the different priorities of the two technical committees even clearer.

GB/Z 218.1 had 31 drafting organizations. The China Electronics Standardization Institute, Beijing Innovation Center of Humanoid Robotics and Beijing University of Technology headed the list, followed by Huawei Cloud, UBTECH, AgiBot, Galbot, Xiaomi Robotics, PaXini, Ant Group, the Zhejiang Institute of Quality Sciences and Beihang University. Robot makers, model developers, standards bodies, universities and research institutes were all represented.

Of TC609’s three projects, “Data Sources and Constituent Elements” has the longest roster, with 17 organizations. The first few are again the Beijing Innovation Center of Humanoid Robotics, PaXini, UBTECH, Xiaomi and Huawei Cloud. Farther down the list, however, are China Mobile (Hangzhou) Information Technology, China Unicom Data Intelligence, Mashang Consumer Finance, the COSMOPlat Industrial Intelligence Research Institute and the Shenzhen Institute of Standards and Technology.

The additional participants are telecom operators, a consumer finance company, an industrial internet platform and a local standards institute. Those who build machines determine whether data is good enough; those who use data define where it comes from and what it contains. The former provides a yardstick for acceptance, while the latter lays the groundwork for circulation. The National Data Administration is focused on the latter: only when a dataset’s provenance is traceable and its constituent elements are clearly defined can it be priced, traded and reused by another party.

The Costly Part Cannot Be Cut

Of the three areas Liu listed, physical-robot collection and human demonstrations concern where data comes from. The third moves up to the interface layer. When a robot performs a task, how should its current state, task instructions, environmental description and action history be organized into context that a model can understand? And how should that context be transferred across different robot embodiments? The industry’s complaint that “every embodiment has its own data system” stems partly from inconsistent collection formats and partly from the lack of a common language at this layer. Officials provided only the phrase “context protocols,” without elaborating.

The economics of the data are already clear. Li Shuai, deputy director of the Artificial Intelligence Center at the China Electronic Product Reliability and Environmental Testing Research Institute, cited a set of figures at this month’s World Internet Conference Wuzhen Summit: the effective stock of embodied AI data is on the order of 1 million hours, while leading vendors already require around 100 million hours for training, a gap of two orders of magnitude. Collecting data through teleoperation of physical robots costs 500 to 1,000 yuan per hour, 100 times the cost of simulation and synthetic-data methods. Yet using just about 10% real-robot data can significantly improve success rates, while an all-real-robot dataset can raise the rate to 76.8%.

The expensive part is precisely what cannot be eliminated. That is the real problem standardization must solve. More than 70 embodied AI training facilities are already operating across China, with over 40 more under construction or in planning, while data platforms have delivered 1.5 million hours of human video data. The faster these facilities proliferate, the more wasteful it becomes for each to collect data under its own standards.