Abstract: For a cheetah-mimicking quadruped robot with 2-DOF (degree of freedom) active spine joints, a multi-model fusion based control approach is presented based on task decomposition principles and biological neural system mechanisms. In this approach, the spring loaded inverted pendulum model is used to realize the single leg jump control, the central pattern generator (CPG) is used to realize coordinated control among four legs and between the spine and legs, and the virtual model control is used to realize the interaction between the robot and environment. A CPG output-based finite state machine is adopted to coordinate the three control models, and a hierarchical multi-model fusion based controller is designed for the cheetah-mimicking quadruped robot. Referring to the motion characteristics of cheetah spine, the motion pattern of robot spine joints is designed, and the coordinated control strategies of the spine and legs are proposed. Finally, a virtual prototype of the cheetah-mimicking quadruped robot is built using Webots. The robot realizes the spine-legs coordinated control in different gaits, runs stably on rough terrains and achieves a smooth trot-gallop-trot gait transition. The simulation results verify the effectiveness of the multi-model fusion based motion control approach of the quadruped robot.