Seventy Graduates, Forty-Six Robots, and the Chinese Factory Where Machines Learn to Fold Laundry

Seventy Graduates, Forty-Six Robots, and the Chinese Factory Where Machines Learn to Fold Laundry

In a 12,000-square-metre facility in Wuhan's East Lake High-tech Development Zone, a young programme manager named Zhang Jia oversees one of the strangest classrooms in the world. His students are not human. They are 46 humanoid robots learning to serve steamed buns, wipe tables, and fold clothes, guided by 70 recent graduates working eight-hour shifts in what China calls a "robot training farm."

The Hubei Humanoid Robot Innovation Centre is one of at least 40 state-funded robot training centres now operating across China, from coastal Hangzhou to inland Mianyang. Together, they represent Beijing's most ambitious attempt yet to solve a problem that has stumped robotics researchers for decades: teaching machines to move through the messy, unpredictable physical world.

How You Teach a Robot to Serve a Steamed Bun

The training process is part repetition, part performance capture, and part brute-force data collection. Human trainers strap on VR headsets and motion-capture controllers, then physically demonstrate tasks while the robots mirror their movements in real time. A single action, such as picking up a plate and placing it on a table, may be repeated hundreds or even thousands of times before the robot can execute it independently.

Cameras and sensors record every movement. Rows of annotators sit at workstations reviewing video footage frame by frame, tagging each moment with labels like "turn left," "extend arm," or "grip object." The Wuhan centre alone produces roughly 100 hours of usable training data every day.

We collect and organise the data, then upload it to our platform, where we classify and process it. But we are still in the exploratory phase.

The facility cost 200 million yuan ($29 million) to build and features laboratory-scale replicas of kitchens, bedrooms, and living spaces where robots practise domestic tasks. It is essentially a full-sized house built for machines to learn in.

The Scale of China's Robot Training Push

What makes the Wuhan centre significant is not the technology itself, which builds on established techniques in imitation learning and teleoperation, but the sheer scale at which China is deploying it. The country's 2026-2030 Five-Year Plan, approved by the top legislature on 14 March 2026, identifies "embodied intelligence" as one of six priority future industries. Beijing is calling for "extraordinary measures" to accelerate the development of humanoid robots, training centres, and the AI models that power them.

Hubei Province alone has committed a 10-billion-yuan government fund to humanoid robot development. Local governments across the country are competing to host training centres, attracted by the prospect of becoming regional hubs for what Beijing expects to be a multi-trillion-yuan industry.

We provide 23 high-fidelity simulation scenarios and over 10 temporary ones, capable of training hundreds of robots simultaneously, collecting more than a million real-machine data points annually. These data, after review, labelling and cleaning, are used for large-model training, enabling continuous robot evolution.

From Data Farms to Real-World Deployment

The data produced in centres like Wuhan feeds directly into the foundation models that power China's growing fleet of humanoid robots. In Sichuan, the Walker S2 humanoid robot is being put through similar paces at a dedicated multimodal data collection centre. In Beijing's western suburbs, another facility uses VR and motion-capture systems to train robots for warehouse operations, material sorting, and product packaging.

The applications extend well beyond domestic chores. Chinese restaurants are already deploying AI-powered robots in commercial kitchens, and the country's short-drama industry is using AI to compress production timelines from months to weeks. The robot training farms represent the physical counterpart to that digital acceleration: teaching machines to do with their bodies what large language models already do with text.

The Workers Behind the Robots

The human side of the story is equally striking. The 70 trainers at the Wuhan centre are mostly recent graduates, young workers who spend their days wearing VR equipment and repeatedly demonstrating physical tasks so that machines can learn from their movements. It is physically demanding, repetitive work, a new kind of labour born from the AI age.

• Trainers work eight-hour shifts guiding robots through task demonstrations
• Each trainer may repeat a single action thousands of times per shift
• Video annotators review footage frame by frame, adding movement labels every few seconds
• The work requires physical stamina, attention to detail, and patience with slow-learning machines
• Most trainers are recent university graduates entering a job category that did not exist two years ago

The irony is not lost on observers: a generation of educated young Chinese workers is being employed to teach robots the manual skills that those robots may eventually perform instead of human workers. A think-tank report forecasts that by 2045, China could have over 100 million humanoid robots deployed across industries, creating a total market value of approximately 10 trillion yuan.

The Global Race for Physical AI

China is not alone in pursuing embodied intelligence, but it is investing at a scale that few countries can match. The combination of government funding, a large pool of young workers willing to do repetitive training tasks, and an existing manufacturing base gives China structural advantages in the physical AI race. President Xi Jinping has framed the push as part of a broader campaign to make China the world's leading superpower in science and technology.

For the rest of Asia, the implications are significant. Countries like Singapore are investing heavily in AI skills, but the robot training farm model requires a different kind of infrastructure: large physical spaces, cheap labour for data collection, and government willingness to fund speculative industrial policy. ASEAN nations building AI governance frameworks will need to decide whether to compete in physical AI or focus on software and services.