To overcome the bottleneck of insufficient human teaching resources for large-scale robot deployment, the team proposed the Collective Experience Learning (CEL) paradigm and completed prototype validation using 96 real robots. This paradigm shifts the human role from demonstrating actions to judging outcomes. Through the Failure-guided Experience Retrieval Fine-Tuning (FERT) algorithm, it uses only sparse feedback on task success or failure to pinpoint error-prone steps and match suitable successful experiences to support learning. In addition, a Fault-Tolerant Asynchronous Training System (FATS) was designed to address issues such as limited communication and local failures in multi-robot collaborative training, ensuring the stable operation of large-scale clusters. In tests across five types of cross-scenario tasks—flower arrangement, clothes folding, mobile phone charging, stool assembly, and mobile phone packaging—new robots that did not participate in training achieved an overall success rate of 97.5%, exceeding the average level of human teleoperators. Moreover, it demonstrated a scaling trend in which, within the same physical training time, the larger the cluster size, the higher the task success rate, offering embodied intelligence a new pathway for continuous growth through autonomous interaction within robot collectives.