Product1 publisher2 min readPublished
China's data regulator will write the training-data standards its robot makers asked for
China's National Data Administration says it will develop standards for embodied AI data, ten days after seven companies asked for exactly that. More than 70 Chinese sites are already recording the hours.
The Product Desk · Product desk
What happened
- China's National Data Administration said it will develop standards for embodied AI data, guide local authorities on the work and support companies putting more money into data resources, Bloomberg reported.
- The pledge followed a 10 September meeting chaired by the agency's head, Liu Liehong, with research institutes, technology companies and humanoid robotics organisations in the room.
- Ten days before that meeting, the same agency heard seven companies ask for public data infrastructure for embodied AI and common data standards, MLex reported.
- More than 70 embodied AI training grounds are already operating across over half of China's provincial regions, with 46 more planned or under construction, according to TechNode.
- Those grounds cluster in three regions: the Yangtze River Delta, the Beijing-Tianjin-Hebei region and the Pearl River Delta.
Compiled by The Product DeskSomething wrong?How this is made
Why it matters
- constraint About 86% of China's training grounds aim at industrial manufacturing, so the corpus these standards would govern is factory work, and a team building for homes or care settings still has to record its own hours.
- exposure European machine builders face Data Act design obligations on connected products placed on the market after 12 September 2026, and whoever receives that data is barred from using it to build a competing product.
- decision With the Data Union Strategy's data labs not set up, a European robotics team deciding where to record demonstrations this year chooses between a commercial gym and a site abroad.
- contradiction TNW reports that only 23% of Chinese enterprises surveyed this year were satisfied with the robots on offer, so the buildout is not yet evidence that anyone has a product customers want.
Before a humanoid robot learns to empty a bin, somebody has to perform the task, recorded, for hours. The China Academy of Information and Communications Technology puts the requirement for one embodied AI foundation model at roughly 10 million hours of real training data [5]. Between 100,000 and 1 million hours of high quality data exist worldwide [6]. Everything recorded so far, by everyone, comes to between 1% and 10% of what a single model is estimated to need [16]. According to TNW, the constraint in humanoid robotics is no longer the model but hours of recorded reality [23].
Recording hours costs payroll and floorspace. Add the 46 grounds under construction or planned to the more than 70 already running and China is on course for at least 116 sites [17]. A common data format is the cheaper of the two things the seven companies asked for, and it is what turns hours recorded at one site into training data another lab can load [3].
Europe wrote the access rules before it built any supply. The Data Act has applied since 12 September 2025 and governs who may reach the data a connected product generates, industrial machinery included [10]. The nearest European equivalent to a Chinese training ground belongs to a company: NEURA Robotics has ten gyms planned and five meant to be running by the end of this year, split between Europe, the United States and China [14]. Those five would be outnumbered by China's operating grounds by at least 14 to one [19].
The report does not say how many hours those grounds have recorded. Divide the 10 million hour estimate across 116 sites and each one owes about 86,000 hours [24], which at a single recording stream per site is close to ten years of round-the-clock work [25].
Two things settle where a robotics team spends next quarter: whether recorded demonstrations of its actual task exist, and whether it can legally reach them and read their format. Yes to both leaves a modelling problem, and TNW's reading is that models are not what is holding this up [23]. Yes to the first and no to the second is a contracts and format problem, which is what the Data Act governs in Europe [10] and what the National Data Administration says it will standardize in China [1]. No to both puts a team in the business of hiring operators and renting floorspace. China's data regulator has just offered to standardize that business and to back companies spending more on data resources [1].
What to watch
- Whether the National Data Administration publishes an actual standard, and whether it covers demonstration data formats or only access and security.
- Whether the Commission carries out the Data Union Strategy action that would set up the data labs, and on what timetable.
- Whether any Chinese training ground reports recorded hours, which would make the 70-site count comparable with the 10 million hour estimate.