An improved collision avoidance behavior of multi-agents was presented due to the fact that the existing methods do not comform to the reality and are devoid of universality. By introducing the role and personality parameters, the agents adjust their collision avoidance behaviors. The improved method constructs a link between collision avoidance behaviors and the attributes of real humans. In this way, the consistency and smoothness of humans' motion is retained. So it fits in with the real humans' collision avoidance behavior more tightly. Also it is more hominine and universal than the existing methods. Simulation tests show that the improved collision avoidance behavior accords with the characteristics of the real people. So it can be used in virtual crowd simulation, multi-robot system and multi-agent system more efficiently.