A robot that can perform a martial arts routine on national television and a robot that can unload a truck are separated by an enormous gap — one that cannot be closed with better motors. It has to be closed with data: millions of examples of real tasks in real places.
In 2026, China started building the infrastructure to collect that data.
Key takeaways
- National programme: MIIT and the state assets regulator launched a 2026 real-scenario training special action for humanoid robots and embodied intelligence, targeting 100+ high-value application scenarios by the end of 2026 and 10,000-unit-scale deployment capability.
- Training fields: Chongqing's Bishan district is building the city's first large-scale embodied intelligence robot training field, with 60 robots in its first phase, covering retail, industry, cultural tourism, eldercare and agriculture — with ten categories of scenario planned.
- Data centres: Guangxi has established an embodied intelligence data collection and testing centre where robots are trained on concrete tasks — including, famously, packaging the region's snail-noodle specialty.
- The bottleneck: as one Chinese commentary put it, robots can already walk, run and dance — but entering factories to work, communities to serve, and greenhouses to harvest is "still a long way off."
Why demonstrations stopped being enough
At the 2026 World Robot Conference, the exhibits were noticeably less theatrical and more vocational. Reports described robots taking on "new job types": table tennis partners and lion dancers, but also housekeepers folding clothes and clearing tables; industrial logistics solutions covering a full chain of high-bay retrieval, flat transport, tote circulation and fine picking; pest-monitoring robots for agriculture; and what was described as the world's first six-in-one AI orthopaedic surgery robot.
International coverage picked up the shift. Reuters framed Chinese robotics as moving "from pleasing the public toward broader commercial use." The BBC noted that Chinese-made warehouse robots are entering British retail and logistics, so that when a UK consumer clicks buy, the item may be moved by a machine built in China.
The Guardian's angle was on dexterity — a crop of Chinese startups using the country's manufacturing base to attack what it called the hardest problem in robotics: a fully dexterous robotic hand.
Training a robot is a data problem
The insight behind the training fields is that general-purpose capability does not emerge from better simulation alone. A robot that has learned to pick a tomato in a greenhouse in Chongqing has learned something specific: how that fruit sits on the vine, how it deforms, what the lighting does.
Bishan's greenhouse deployment illustrates the loop. A robot dog equipped with HD vision, thermal imaging and lidar patrols tomato rows, analysing leaf condition, identifying early disease and measuring fruit diameter, then feeds that into the greenhouse control system — quality fruit rate reported above 90%. A harvesting robot works alongside it. The training field is where the next generation of that behaviour gets learned.
Guangxi's centre does the same for industrial and logistics tasks, using VR to capture human operators performing real work and converting it into training data.
What "AI Plus" adds
The push is policy-backed from the top. China's 2026 Government Work Report made "deepening and expanding AI Plus" an annual priority, and the State Council's Opinions on Deeply Implementing the "AI Plus" Initiative set staged targets: deep integration with six key areas and over 70% adoption of next-generation smart terminals and agents by 2027, over 90% by 2030, and a fully intelligent economy and society by 2035.
For robotics specifically, the 15th Five-Year Plan outline names embodied intelligence as a future industry. The practical translation is that scenario owners — factories, warehouses, hospitals, farms — are being asked to open their doors as training venues.
The honest state of play
It would be a mistake to read 18,000 units as 18,000 working employees. Customer structure is still concentrated in education, research and exhibition. Industrial deployment is at pilot scale. Autonomy software remains immature relative to hardware, and early-generation humanoids require frequent calibration — a common trait that Chinese vendors are trying to engineer out through manufacturing consistency rather than software cleverness.
But the direction of investment has changed. The money is no longer going into making robots more impressive; it is going into making them more useful. That is a much harder problem, and China is attacking it the way it attacks manufacturing problems — with volume, data and infrastructure.
Based on MIIT and State Council policy documents, World Robot Conference 2026 reporting, and Chinese and international media coverage in 2026.