China's 70+ Robot Training Schools Struggle to Sustain Without Data Sales
In training grounds across China, rows of embodied AI robots practice tasks under remote control by data trainers—grasping water bottles, placing meal trays in microwaves, scanning packages. Nearby, operators wearing motion capture gear fold clothes and make beds, every movement recorded precisely by sensors to become teaching material for robots.
Such scenes are unfolding in embodied intelligence training grounds nationwide. A recent report by the China Academy of Information and Communications Technology (CAICT), the "Embodied Intelligence Training Ground Research Report (2026)," shows that by the end of June this year, over 70 training grounds had been built and put into operation, with more than 40 under construction or planned.
The distribution of training grounds follows application scenarios rather than city tiers. It closely aligns with China's embodied intelligence industrial clusters, forming three major clusters centered on the Yangtze River Delta, Beijing-Tianjin-Hebei, and Pearl River Delta. Zhejiang has the most built training grounds at 14; Jiangsu, Beijing, Guangdong, and Shandong follow with 8 each.
Zhang Weimin, deputy director of the Embodied Intelligence Department at CAICT's AI Research Institute, notes a notable trend: training ground construction is breaking away from the traditional AI industry's concentration in first-tier cities, extending to third-, fourth-, and even fifth-tier cities, showing a multi-level city coordination layout. This shift reflects a move from talent- and capital-driven to real-scenario, data-supply, and operational-efficiency-driven construction—wherever there is a suitable scenario, a training ground is built.
Training grounds are not just data collection sites. Gao Fang, head of the Guangdong Embodied Intelligence Training Ground, calls it a "robot school." The logic: robots graduate from school and need to find suitable jobs. The Greater Bay Area has rich scenarios and robot supply chain players. The Guangdong training ground regularly facilitates precise matchmaking between robot products and industry application side, helping robot companies find real deployment scenarios and resources for robot brain training, and helping manufacturing, healthcare, energy and other industries find suitable embodied AI robot solutions.
The National AI Application Pilot Base in Hangzhou, Zhejiang, acts as a connector. With a mixed-ownership enterprise as the construction entity, it builds an open cooperation mechanism led by state-owned enterprises with participation of industry leaders and application scenario enterprises. Zhang Haiwei, founder of Qingtong Robot, notes that in the past, developing an application scenario cost tens of millions just for training data; now, by settling in the base, companies can access needed resources in one stop.
Standardization is just starting, and the business model is still being explored. Standardization is key to improving training effectiveness. An industry standard on "Quality Requirements and Evaluation Methods for Embodied Intelligence Datasets," drafted by CAICT with over 40 units, will be officially implemented on November 1. Zhang Weimin says the release marks a shift from scale-oriented to quality-oriented dataset construction, filling a gap in industry standards for embodied intelligence data.
But training grounds face more realistic challenges. Most rely on selling data products as their main revenue source, which alone is hard to cover investment. Wang Maolin, co-founder of Shenzhen Jubaopen, says the next step is to transition from single data supply to training-as-a-service covering data production, model training, and application validation, building a diversified revenue structure and moving from heavy asset investment to sustainable operation.
Jiang Lei, chief scientist of the National and Local Joint Humanoid Robot Innovation Center, judges that embodied AI robots have achieved 0-to-1 technological breakthroughs, but the 1-to-10 scale deployment problem remains unsolved, stuck in standard systems, engineering management, and pilot testing, with the core contradiction concentrated in scenario adaptation. Building training grounds is only the first step; whether they can operate sustainably depends on finding a path to self-sustain among data, models, and scenario validation.