https://j.people.com.cn/n3/2026/0821/c95952-20491104.html
Over 70 Embodied AI Training Facilities Operational; Industrial Implementation Accelerates
According to the China Academy of Information and Communications Technology (CAICT) “Embodied AI Training Facility Research Report (2026)”, over 70 embodied AI training facilities are currently completed and operational nationwide, with more than 40 additional facilities under construction or in the planning stages. Leveraging the strengths of the embodied AI industry, these facilities have formed three major core clusters
- the Yangtze River Delta,
- Beijing-Tianjin-Hebei, and
- the Pearl River Delta
and cover more than half of China’s provincial-level administrative regions.
Experts note that in terms of embodied AI applications, highly repetitive and modular tasks—such as assembly and manufacturing, component feeding and transfer, and material transport—are expected to complete training first and see industrial implementation relatively early. Conversely, industrial implementation in lifestyle service sectors involving human interaction will likely lag behind due to considerations such as AI ethical standards.
A visit to an embodied AI company in Beijing’s Zhongguancun district revealed that its data platform aggregates human video data collected from across the country alongside synthetically generated simulation data. According to a company executive, the firm has already delivered 1.5 million hours of human video data for training purposes—a volume that ranks among the highest globally. However, training high-quality embodied large language models often requires datasets exceeding 10 million or even 100 million hours. This immense market demand is driving the company to continuously refine its embodied AI training and evaluation processes.
Wheeled, arm-equipped humanoid robots designed to operate power distribution panels have been trained on vast amounts of real-world operational data. They can now accurately identify control buttons of varying colors and positions, and autonomously perform complex, high-risk tasks such as flipping switches and opening or closing industrial valves. Several power inspection robots have already completed algorithmic tuning and are currently being delivered.
According to Dong Dianbiao, an executive at the robotics company, through training on tens of thousands of data samples, robots have achieved an 80–90% success rate in their current tasks. Through incorporation of additional data, the success rate will comecloser to 100%.