On August 21 we covered the 70-plus embodied-AI training grounds China has already built, 86% of them laid out as industrial manufacturing. The open question then was who writes the spec for what gets recorded inside them. That line moved this month, and industry asked first.
On September 13 Bloomberg reported that China’s National Data Administration will develop standards for embodied-AI data and guide local authorities on the work. Ten days earlier the same agency had met seven companies requesting public data infrastructure and common data standards, per MLex. Ten days from ask to answer.
Who was at the table
The agency’s own readout puts the September 10 symposium under NDA director Liu Liehong, with deputy director Xia Bing attending (carried by Jiemian). Nine organisations spoke on the theme of data enabling embodied intelligence:
the Chinese Academy of Sciences Institute of Automation, the Beijing Institute for General Artificial Intelligence, Lightwheel, JD.com, Deep Robotics Intelligence, the Shijingshan Humanoid Robot Data Training Centre, Galaxea, Zi Bian Liang, and the Beijing Humanoid Robot Innovation Centre.
Read that list as an org chart. Research institutes, robot builders, a synthetic-data firm, JD.com holding real warehouse and delivery environments, and one organisation named for the job itself: a data training centre. Collecting the hours has become its own institution rather than a department inside a robot company.
The agency committed to three things. Develop embodied-AI data standards in due course, guide provincial and municipal data authorities, support companies putting more money into data resources. No draft, no deadline, no consultation window.
The size of the hole
One gap explains the urgency.
CAICT estimates embodied-AI foundation models need roughly 10 million hours of real training data. Between 100,000 and 1 million hours of high-quality data exist worldwide. That is one to two orders of magnitude short.
The language-model build-out is the wrong cohort, and the reason it is wrong is the point. Text was already sitting there. Decades of it, marginal collection cost near zero, the remaining work being cleaning and licensing. Embodied data has no such inventory. Ten million hours is not a download. Somebody has to perform those hours at a real workstation and record them, one hour at a time. That is why a training ground has to be a building, and why the bottleneck in humanoid robotics has never been the model.
So the four things the standards will govern are collection, labelling, storage and transfer. Put that on a shop floor: the feel a worker built at one station over several years, recorded to what spec, labelled under whose scheme, held by whom, sold to whom.
Europe wrote the rules first and has nothing to apply them to
The sequence is inverted in Europe.
The Data Act has applied since 12 September 2025. It governs who may reach the data a connected product generates, industrial machinery included, and it forbids using that data to build a competing product. Design obligations bite on connected products placed on the EU market after 12 September 2026. The rules are more specific than anything China has published.
What is missing is the supply. The Data Union Strategy of 19 November 2025 opens with scaling access to data for AI, data labs included. Ten months on, most of those actions have not been carried out. Europe’s nearest equivalent to a Chinese training ground is a single company: NEURA Robotics is building gyms, ten planned and five meant to be running by the end of this year, split across Europe, the United States and China (TNW).
Collect first and regulate after, or regulate first and collect later. Both orders carry a cost. Only the first one is moving a timetable for when a machine can hold down a real workstation.
Where a data standard actually lands
The word “standard” sounds far from the factory floor. It is not.
A collection spec settles what a recorded motion has to look like to count: how many camera positions, what sampling rate, whether force feedback is captured, whether the failed attempt is kept alongside the good one. Those numbers decide how people work inside a training ground. As we noted in the August piece, most of the people producing that data are experienced line workers pulled off production, chosen precisely because the skill is in their hands. Once the spec exists, that work stops being “help R&D record some material” and becomes a role with acceptance criteria: hours logged, pass rate, what counts as a redo.
Labelling and storage standards land on a different group. Somebody decides this grasp succeeded and that one slipped, somebody clears the bad samples. That is existing labour, and it scales with the hour count.
The transfer standard is the one that touches workers most directly, and it is the vaguest line in the readout. Operating data recorded at one station: who owns it, whether it can be sold, to whom, and whether the same capture can train a robot for a different factory. Nobody at the table speaks to that. Among the nine organisations that presented, none is a union or a labour-relations body. The worker is the party being recorded, not the party negotiating the spec.
Do not read this as settled
A statement is not a document. No draft, no consultation text, no effective date, and “in due course” leaves a lot of room.
Demand has not caught up either. Only 23% of Chinese enterprises surveyed this year were satisfied with the robots on offer. A training ground banks hours, not orders. A standard lowers the cost of exchanging data; it does not raise the rate at which a robot finishes a shift on a live line.
Two dates worth watching. First, when a draft standard opens for public comment — that is when the collection spec’s precision becomes visible, and when the transfer clause gets written down. Second, when the first dataset accepted under the new standard is delivered. Our August window for capability reaching real plants was around 2027, concentrated in the Yangtze River Delta, Beijing-Tianjin-Hebei and the Pearl River Delta. This week does not move that window. It changes whether the data inside it travels between plants.
Sources
- NDA director Liu Liehong chairs embodied-intelligence symposium — Jiemian
- China’s data regulator plans standards push for embodied AI — Bloomberg
- China’s data regulator plans standards for embodied AI, ten days after industry asked — The Next Web
- China to support embodied AI companies’ data investments, promote standards — MLex
- Data Act — European Commission
