Building an Embodied AI Training Base: Three Procurement Tiers from RMB 4.4m to RMB 77.8m

Embodied AI training bases are among the fastest-growing procurement categories of 2026. In the first half of the year China awarded 218 humanoid robot contracts worth over RMB 1.723 billion, up 163% in count and 422% in value year on year. Education and research institutions accounted for 55.5% of the projects, while government and enterprises together made up 41.7% and drove most of the value.

The market is highly concentrated

Projects above RMB 10 million are only 8.7% of the count but contribute nearly 90% of the value, and five billion-yuan projects account for more than half of the total. Large awards go to embodied AI innovation centres, data-collection bases and industry service platforms, which means buyers are moving from purchasing devices to purchasing platforms. Demand has split into three lines: teaching and research equipment, government industry platforms, and industrial scenario validation.

Three base formats and their budgets

Colleges: workstations matter more than robot count

The key purchase in a college project is not how many robots but how many workstations. The Beijing technician college notice spells them out: real-machine data-collection workstations for humanoid, wheeled and robotic-arm forms, plus joint-motor and whole-body assembly test workstations. Workstation count follows how many students are in the room; form count follows course coverage. Those two determine whether the base can actually teach, while robot count only determines how good the showroom looks.

Regional centres: public money buying data capacity

Regional data-collection centres are usually funded by local governments or state-owned platforms. The model: the government funds the centre, buys robots in bulk, runs data collection in purpose-built scenarios, and the resulting datasets serve robot companies and algorithm teams. By the end of April 2026, 64 data-collection and training centres were operating in China, more than 90 including those under construction, with over 2,500 robots deployed across more than 280,000 square metres. Shanghai-based AgiBot leads with about 200 units; centres in Zigong, Huizhou and Jiangsu have reached 150 or more.

Why local governments pay for this

Because data is the scarcest resource in the industry. Commercial deployment is generally estimated to need tens of millions of hours of real-world operating data, while the entire industry has only a few hundred thousand hours, a gap of two orders of magnitude. Collection is slow: with teleoperation, one operator hour yields about one hour of valid data. IDC data shows that in 2025 shipments of humanoid robots went 78.4% to performances, research and data collection, with less than 20% into real factory work. Policy is adding fuel: Jiangsu has committed to opening at least 100 real application scenarios a year with up to RMB 1 million of support per project, and the national dataset management service platform is in trial operation with 516 certified bodies and 1,350 published datasets.

Four decisions before project approval

1. Positioning: teaching, regional data capacity, or industry services? The equipment structure and acceptance criteria differ completely
2. Scenario list: fix scenarios and courses first, then forms and unit counts
3. Data ownership: who owns collected data and trained models must be in the contract
4. Operations: who staffs it, who maintains it, who buys the data. Without an operating plan the base goes idle in year two

Settling these four decisions before approval is worth more than adding one more robot to the budget. Bases usually fail on the operating model, not the equipment list.