China Builds Robot Training Grounds to Fix Data Gap
China has opened over 70 facilities to train humanoid robots on real-world tasks, addressing a critical shortage of physical interaction data.
Key points
- China has opened over 70 embodied AI training grounds to generate the physical interaction data robots need for real-world tasks.
- These facilities reduce development costs by allowing multiple companies to share testing environments and data infrastructure in key economic zones.
- A new industry standard taking effect in November aims to improve the quality of training datasets, moving beyond simple volume accumulation.
China is rapidly expanding a network of specialized facilities designed to train humanoid robots on real-world tasks. These "embodied AI training grounds" serve as a crucial infrastructure layer, addressing a fundamental bottleneck in the industry: the lack of systematic data on how robots interact with unpredictable physical environments.
While the country has shipped over 14,000 humanoid units in 2025, accounting for the majority of global output, these machines often struggle to move beyond controlled demonstrations. The new training grounds aim to bridge the gap between laboratory performance and reliable everyday utility by providing the repetitive, data-intensive practice that digital models do not.
Data scarcity limits robot capability
Large language models learned from the vast text available on the internet, but robots lack a similar digital reservoir for physical actions. According to macaubusiness.com, physical-world interaction data has never been collected in a standardized way. This means developers cannot simply download a dataset; they must generate it through repeated, supervised trials in real spaces, a process that is slow and expensive without dedicated infrastructure.
Regional hubs create shared ecosystems
As of mid-2026, more than 70 of these facilities are operational in China, with another 40 under construction. They are concentrated in key economic zones like the Yangtze River Delta and the Pearl River Delta. In Guangdong, a provincial facility acts as a "robot school," connecting manufacturers with healthcare and energy providers to test solutions in actual working environments.
This centralized approach reduces costs significantly. A founder of a Hangzhou-based robot firm noted that previously developing a single application scenario could cost tens of millions of yuan in data acquisition. By moving into shared bases like the National Pilot Base for Embodied AI Applications, companies can access diverse testing scenarios and necessary infrastructure in one location, turning isolated efforts into a collaborative ecosystem.
Policy mandates standardized data quality
Government support is accelerating this shift. In June 2026, authorities launched a special action plan requiring each province to identify at least 20 priority scenarios for real-scene training. This policy push is paired with an industry standard on dataset quality, drafted by over 40 organizations, which takes effect in November. The standard aims to shift the focus from merely accumulating data volume to ensuring the data is high-quality and reliable, addressing a long-standing gap in the field.
