Chang Lin, CEO of Leju, said at the conference that real-robot data is core to scaling embodied AI. He introduced a training path of pretraining, mid-training, post-training, and reinforcement on real machines, stressing that data reuse determines scalability. Leju has achieved routine operation of about 15 robots in an auto parts loading scenario at Hangzhou Zhaofeng, reaching 60%-70% of human efficiency. Through the OpenLET community, Leju has open-sourced 100,000 minutes of real-robot data with over 6 million downloads.
On October 15, Leju and Zhaofeng jointly launched an embodied AI solution for automotive parts loading at Zhaofeng’s Hangzhou factory, with 15 humanoid robots now in routine operation across 15 stations handling more than 10 types of parts. Each loading cycle takes under 40 seconds with a success rate of about 98%; the project has been running for over two months, accumulating about 1,000 hours and helping produce more than 60,000 parts. The two companies also donated the real-robot loading dataset to the OpenLET community of the OpenAtom Open Source Foundation.
At a media preview of the 26th China International Industry Fair on October 11, Leju Robotics announced it will debut a scenario-based solution centered on a universal humanoid body and universal dexterous hands, covering research, commercial service, and industrial sectors. The mature deliverable solutions include the Kuafu robot guide and Taskor industrial deployment, with industrial solutions already ordered by over 10 leading customers including FAW and Hisense.