The only Chinese team moving a 10,000-GPU autonomous-driving organisation wholesale to robots
The only Chinese team to move an intact autonomous-driving organisation — trained on tens of millions of hours of data, ten-thousand-GPU infrastructure and million-vehicle mass production — into robotics, building a real-time, scalable physical-native foundation model.
Chuguang Yueqian takes a team that has already industrialised end-to-end models in autonomous driving — tens of millions of training hours, ten-thousand-card clusters, millions of production vehicles — and applies the same scaling system to robots. It combines a World Mind Lab world-model team, a North American humanoid-control team and robot hardware and commercialisation teams, starting from 'seeing the physical world clearly' to build a physical-native foundation model that runs in real time and keeps scaling. Its first industrial deployment is on precision optical-module production lines with Zhongji Innolight.
A scalable 'fast system' handling high-frequency physical perception, world prediction, action generation and real-time correction, paired with a 'slow system' that takes semantic understanding and long-horizon planning from general large models.
The team's infrastructure delivered 2× engineering training efficiency, 3× from high-value data mining, 5× overall training efficiency, storage and iteration cost cut to one third, and model iteration cycles reduced from weeks to days.
Long-horizon manipulation tasks on precision manufacturing lines, starting with optical-module production at Zhongji Innolight, developed as an industry benchmark.
Other companies in this category
No. 1 in China’s 2025 outdoor cleaning robot shipments, operating at 100+ sites
China's first ten-billion-RMB world-model unicorn — three global benchmark firsts