With the more robots, the information interaction of multi-robot system becomes more sophisticated and important in a community perception network environment. By exploiting and fusing the learning information of robots in a perception community, the community information sharing mechanism is proposed, as well as updating rules of the community Q-value table. Moreover, considering the existence of delays of learning information transmission, an improved Q-learning method based on homogeneous delays is presented to improve the robot learning efficiency over the community perception network. Finally, the test experiments demonstrate the effectiveness of the proposed scheme.