Zhiyuan, with Tsinghua University and Wuwen Xinqun, released a Collective
Experience Learning (CEL) paradigm using a FERT failure-guided
experience-retrieval algorithm and a FATS fault-tolerant asynchronous training
system. A 96-robot physical cluster demonstrated “robot teaches robot,” reducing
dependence on human expert demonstrations. In five real-world tasks including
flower arranging, clothes folding and phone packaging, CEL outperformed
conventional training methods, with peak overall success rates up to 16.7
percentage points above the average of human remote operators. With equal
training time, larger cluster scale increased operational reliability. CEL is
presented as addressing the industry bottleneck of insufficient human
Demonstration supply for large-scale embodied intelligence deployment.
