A vocational college in eastern China keeps its robots in a room that used to be a welding workshop. The machines stand on steel benches, waiting to be switched on.
The students are not learning to design a humanoid. They are learning to install one, calibrate its joints, log failures and hand it to a maintenance team that has never seen one. Until September, that job had no official name. Now it does.
Key takeaways
- On 9 September 2026 the human resources ministry published its eighth batch of new occupations: 11 new occupations and 23 new job types (工种) inside existing ones. Two matter here — embodied-intelligence robot application technician (具身智能机器人应用技术员), and agent developer (智能体开发员), which is a job type rather than an occupation (MOHRSS).
- The 2026 undergraduate catalogue added the interdisciplinary category (交叉学科) and, inside it, majors in embodied intelligence and brain-computer science (MOE, April 2026).
- Zhejiang University, one of nine institutions approved in the first round, opened an embodied-intelligence undergraduate programme in May 2026, led by its control-science school and drawing on artificial intelligence, computer science and aerospace.
- MIIT and SASAC want 10,000-unit-scale deployment capability and more than 100 high-value scenarios by the end of 2026.
- Unitree still earned more than 73% of its humanoid revenue from research and education in the first nine months of 2025, and 9% from industrial use.
What an occupation listing actually does
China's occupational catalogue is published by the human resources ministry, and recognition is not a licence. Nobody needs a certificate to bolt a robot to a floor.
What recognition changes is coordination. A named occupation can enter vocational curricula, be given a skill grade, be counted in labour statistics and be bundled into training subsidies. Employers can advertise a job a school already knows how to teach.
The September announcement matters more than its size suggests. Alongside the robot technician, the same batch named a job type for people who build and supervise AI agents — work that exists almost entirely inside software. Both describe work firms already do without a shared vocabulary.
A July 2026 implementing opinion from four ministries — human resources, the NDRC, industry and the data administration — pointed the same way: build an AI employment-impact assessment system and expand skills training and assessment (MOHRSS, 8 July 2026). The instruments are administrative; the intent is industrial.
The bottleneck is people, not motors
The hardware side is moving quickly. TrendForce expects China's humanoid output to rise 94% in 2026, with Unitree and AgiBot taking nearly 80% of global shipments (TrendForce, 9 April 2026). UBTech's Liuzhou plant, in production since September, is designed for more than 10,000 industrial humanoids a year.
Deployment is slower. A MIIT and SASAC programme launched on 9 June 2026 targets 10,000-unit-scale deployment and more than 100 high-value scenarios by year-end, drawing on more than 56,000 basic-level, 13,000 advanced-level and 500 excellence-level smart factories, plus 15 lighthouse-level (领航级) plants where AI reaches more than 70% of business scenarios (MIIT, first half of 2026).
A lighthouse plant can absorb a new machine because it employs people who debug integration for a living. A tier-three supplier cannot. The gap between them is measured in technicians — people who can commission a cell, diagnose a torque fault, retrain a policy when the part changes and decide when to stop the machine.
Three job families the list is really creating
- Commissioning and maintenance. Installing a humanoid, mapping its workspace, setting safety envelopes and keeping it running. The largest near-term demand, and the least glamorous.
- Teleoperation and data collection. Operating robots remotely to generate the physical interaction data that training depends on — already an industry in its own right.
- Supervision and compliance. Deciding when a machine may act alone, keeping logs, answering for an incident. The agent-related titles point here.
None of these requires a frontier researcher. All of them require someone with a recognised qualification, because factories buy assurance as much as they buy arms.
The campus pipeline
Approval is the easy part. The 2026 catalogue gives these degrees a category to sit in, and Zhejiang has already filled it.
Proximity helps. Beijing's E-Town hosts more than 300 robotics firms and a 4,000-square-metre robot retail and service centre carrying more than 100 products from 150-plus channel partners; Shenzhen's Robot Valley in Nanshan gathers more than 200 enterprises across components and applications (CGTN, 15 October 2026).
The bet is plain: build training capacity near the factories, standardise the titles, let the labour market fill in demand.
The risk: credentials before demand
Unitree, the country's most visible humanoid maker, drew more than 73% of its humanoid revenue from research and education in the first nine months of 2025 and just 9% from industrial use (company disclosures). The first paying customers are the institutions that teach people how to use them.
That is how most capital goods begin. But these majors are training students for a market that is still mostly a training market. If deployment lags enrolment, the result is credential inflation: certificates signalling commitment rather than competence.
The care sector is the obvious pool. China had 310 million people aged 60 or over at the end of 2024, about 22% of the population (National Bureau of Statistics). Robots for that market must be boring, slow and safe, and need carers to supervise them — a job with no certification pathway in most provinces.
Honest limitations
Occupation recognition is administrative, and no employment, wage or enrolment figures have been published for the new titles, so the size of the job market is unknown. Curriculum details come from university announcements, not independent evaluation of teaching quality. Deployment targets are policy goals, not outcomes. Unitree's revenue split covers one firm, and the claim that skills now bind deployment is an inference, not a measured finding.
What readers can do now
- Operators: when evaluating a Chinese robot supplier, ask what commissioning and maintenance package ships with the unit, and who is certified to service it in your country; that support layer, not the spec sheet, decides uptime.
- Investors: watch the ratio of education and research revenue to industrial revenue at listed robot makers; a widening gap means the commercial market is still being deferred.
- Researchers: study what the new curricula actually teach and whether graduates can complete tasks unsupervised; labour data will arrive years before productivity data.