Abstract: Research on walking stability of humanoid robot is a hot spot of robot field, gait planning based on dynamics model is the main method, but the motion pattern based on gait planning need complex calculation and the form is single. To generate diverse gait of robot, research on the gait of humanoid robot within a framework of imitation learning was launched, human walking information was used as teaching data to realize the imitation of human. The motion planning of humanoid robot was simplified and more diverse and personate behaviors. In order to satisfy the stability of walking process, the cog deviation was compensated based on zero moment point (ZMP) criterion and the switching of the supporting feet in the walking process was determined by using the hysteresis curve to ensure the stability of robot. The simulation of imitation system based on robot NAO indicates that the utilizatin of ZMP criterion ensure the stability in the process of human gait imitation, and the switching of supporting feet based on hysteresis curve ensure the stability of switching supporting foot.