By the end of June 2026 China had more than 70 embodied-AI training grounds running, with over 40 further sites either being built or in the planning stage, a China Academy of Information and Communications Technology report cited by Economic Information Daily and Xinhua on 22 September shows.
Key points
- 70-plus training grounds operational by end-June 2026; 40-plus more planned or under construction
- Sites concentrate in Yangtze River Delta, Beijing-Tianjin-Hebei and Pearl River Delta; Zhejiang leads with 14
- MIIT and SASAC ask each provincial region to identify at least 20 priority scenarios for humanoid robots and embodied AI
- Industry standard on embodied-AI dataset quality, drafted by CAICT with 40-plus organisations, takes effect 1 November 2026
- Operators testing training-as-a-service models combining data collection, model runs and application validation
Training grounds concentrate in three clusters
The CAICT tally shows Zhejiang hosting 14 sites, with Jiangsu, Beijing, Guangdong and Shandong at eight each. Facilities stage repetitive tasks — grasping bottles, loading trays, scanning parcels — while motion-capture operators generate labelled trajectories. The Guangdong provincial facility also serves as a matchmaking hub linking robot vendors with healthcare and energy users, while Hangzhou’s national pilot base acts as a shared connector for state-owned enterprises, technology firms and application partners.
MIIT and SASAC set provincial scenario quotas
In June 2026 the Ministry of Industry and Information Technology and the State-owned Assets Supervision and Administration Commission launched a special action covering physical-world training for humanoid robots and embodied AI. The special action asks every provincial region to identify at least 20 priority scenarios and to build trainable, testable spaces with minimal retrofit.
CAICT dataset standard takes effect November 1
An industry standard on embodied-AI dataset quality, drafted by CAICT with more than 40 organisations, is scheduled to take effect on 1 November 2026. The standard aims to shift datasets from scale-first collection toward quality and evaluation methods, establishing the first national quality bar for physical AI data.
Operators are also piloting training-as-a-service models that combine data collection, model runs and application validation.
.jpg){kind=link}