The shift begins with a glove calibration. The operator stands inside a frame of cameras, lifts a plastic crate from a low shelf, sets it on a table, and walks back. Then the crate goes back on the shelf, and the sequence repeats.
The crate is empty. The shelf is a prop. What the room produces is not a shipment but a record: the same motion, repeated until it becomes a dataset.
This is the least photographed part of China's robotics build-out, and it may decide whether the rest of it works.
Key takeaways
- The binding input in embodied intelligence (具身智能) is recorded physical interaction, not motors or model weights.
- MIIT and SASAC's June 2026 programme targets 10,000-unit-scale deployment and more than 100 scenarios by the end of 2026.
- China's embodied intelligence market is estimated at RMB 1.09 trillion (≈ US$153bn / HK$1.2tn) in 2026, from RMB 213.3 billion in 2018 (industry report).
- Only 9% of Unitree's humanoid (人形机器人) revenue is industrial; the rest is research, education and demonstration.
- The number of people collecting this data in China, and what they are paid, is not disclosed by any ministry or company here.
Why pixels are not enough
Robotics borrowed its training method from vision, and the loan has not fully cleared. Internet images taught models to recognise; the assumption was that video would teach them to act.
Video shows the result of an action, not its parameters: the force applied, the moment an object slipped, the small correction a hand makes without reporting it. An embodied model needs tuples of observation, action and outcome, and the action half is rarely recorded by a camera pointed at the world.
That data exists in kitchens and warehouses, but it is not indexed, and almost none of it is labelled with what a human's joints actually did. So it has to be manufactured, in rooms like the one above.
How the collection is organised
Four methods do most of the work:
- Teleoperation, where a human moves a leader arm and the follower records joint trajectories.
- Motion capture, where a suited operator performs the task and the skeleton is retargeted onto the machine.
- Mock sites, where factories, kitchens and supermarket aisles are rebuilt at training centres so the environment stays repeatable.
- Live pilots, where real sites open under the MIIT scenario programme and the data arrives dirtier and more valuable.
The economics favour density. A collection site inside an industrial park can draw on the same suppliers, the same labour pool and the same technicians as the factories it imitates, and it can run shifts.
Policy supplies the venues. MIIT and SASAC's programme, launched in June 2026, asks for 10,000-unit-scale deployment capability and more than 100 high-value scenarios by the end of this year (SCIO, 4 August 2026). Every scenario on that list is also a place to record.
The labour question nobody has priced
The operator is the product. Their hands are the training signal, and the recording captures not only the motion but the room, the voice, sometimes the face.
- Consent: no disclosure here describes what a demonstrator agrees to, or for how long their motion can be reused.
- Attribution: a dataset built from one skilled worker can train a model that replaces that class of work, with no payment after the contract ends.
- Strain: repeating one motion for a shift is a known industrial injury pattern, and the work is paced by a camera rather than by a line.
- Privacy: mock sites are easier to control than live ones, where a camera in a ward or a workshop records people who signed nothing.
No public count exists of how many people do this work in China. No ministry or company in these disclosures publishes a headcount, a wage or an injury rate for it. Any specific figure you see is an estimate by whoever printed it.
Volume is not diversity
The temptation in any state-backed collection programme is to count hours, because hours are countable. Ten thousand demonstrations of one crate in one room produce a model that is confident inside that room and brittle one door away.
Two further problems follow. The first is taxonomy: a dataset is only as good as its labels, and "pick", "place" and "hand over" hide the distinctions that decide whether a grasp holds. The second is simulation, which is cheap and wrong in specific ways that are hard to audit; a policy trained mostly in simulation tends to fail on surfaces it has never touched.
China's collection advantage is real, but it is an advantage in volume and cost, not automatically in coverage.
What would make it defensible
- Provenance: who performed each motion, under what consent, and for which permitted uses.
- A published task and failure taxonomy, so one dataset can be compared with another.
- Independent audit before data enters a government-backed training centre.
- Some compensation for the demonstrator that survives the contract, even if thin.
None of this is exotic. These are the conditions under which a data business becomes an asset rather than a liability, and the country that writes them first gets to sell training data abroad without a consent scandal attached.
Honest limitations
The market figure comes from an industry report with a broad definition and is not a measure of spending on training data. Unitree's 9% industrial share is company-disclosed and covers one vendor. The argument that density is China's advantage is an inference from the design of the MIIT programme, not a measured productivity comparison. Worker numbers, wages, dataset sizes and consent practices are undisclosed, and this column has not visited a collection site.
What readers can do now
- Operators: ask any vendor how much of its training data came from your sector, and whether you can inspect the failure taxonomy.
- Investors: treat demonstration hours as a cost line, not a moat, until provenance and reuse rights are documented.
- Researchers: publish consent and provenance protocols alongside datasets; the weak point in this field is documentation, not capture.