Bernstein: 42% of China's 20K+ humanoid robot shipments in 2025 were for education and R&D; data collection and human-robot interaction services were 19% each

Network of training farms aims to supply data needed to put ‘brains’ into machines — At a new 12,000 square metre facility in Wuhan …

Financial Times 2026-03-12 Ryan McMorrow

Context & Ripple Effects

China’s humanoid push has been framed as a contest to build useful embodied AI, backed by a state investment priority for embodied AI. This shipment breakdown adds a practical detail: current deployments are concentrated in learning, experimentation and the collection of training inputs rather than a single mature commercial use case.

The Wuhan training-farm facility fits a sector that related coverage describes as having roughly 140 companies pursuing humanoids. It highlights that scaling robot fleets also requires organized environments in which machines can generate data and improve human interaction.

First-order effects

Education and R&D account for the largest disclosed share of China’s 2025 humanoid shipments, making research and training environments the most immediate deployment channel in Bernstein’s breakdown.
Data collection and human-robot interaction services each represent meaningful deployment categories, while the new Wuhan facility expands capacity explicitly aimed at producing training data for robot AI systems.

Second-order effects

Humanoid makers seeking near-term deployments will have an incentive to compete for access to research sites, data-collection operations and interaction-service partners, not only for end-user hardware sales.
Training farms and similar controlled environments become an adjacent input market: they can turn deployed robots into sources of data needed to refine models and interaction behavior.

Third-order effects

If this mix persists, humanoid competition may be shaped as much by access to data-generating deployment networks as by mechanical design, reinforcing the embodied-AI development race.
The sector could evolve from isolated robot sales toward an integrated stack of hardware, training infrastructure and services; whether that produces broad commercial adoption depends on these early deployments translating into reliable capabilities.

The trend: Humanoid robotics is moving toward data-intensive industrialization, where early fleets are deployed to build the training and interaction feedback loops needed for wider use.

Discussion

@rwmcmorrow Ryan McMorrow on x
The embodied AI data collection project officially got Beijing's blessing last week and was written into the next 5-year plan. It marks a new era for China where its technology ambitions are not just to catch up with the west, but are now actually leading in frontier research [im…
@rwmcmorrow Ryan McMorrow on x
While it's unclear for now if the data collection labs will lead to physical AI breakthroughs or be a waste of money, they are at least providing a market for China's legions of humanoid startups: [image]
@rwmcmorrow Ryan McMorrow on x
I visited some of China's state funded robot training labs which are popping up around the country to collect training data for physical AI Young grads at the labs spend their days showing the robots how to do the same thing over and over again, like picking up packages: [video]
@rwmcmorrow Ryan McMorrow on x
The idea is that the succession of sensor readings and videos that they capture while training the robots can be fed into AI models Eventually it could lead to machines learning more generalised skills, such as picking up a bottle of water without being explicitly programmed to
@nicholasdwright Nicholas Wright on x
Training humanoid robots - key for China US competition as I described last year in @ForeignPolicy https://foreignpolicy.com/... Interesting new developments in @FT by @rwmcmorrow https://www.ft.com/... Inside the labs training China's humanoid robots [image]

Chronicles