We propose a new computational model describing the motion imitation by humans, and which parallels the direct matching hypothesis in Neuroscience. To illustrate our computational model, we present a scheme where a robot imitates another robot achieving stepping motions on horizontal plane. The idea is to map an observed action onto the robot's motor representation of the same action: the motion controller. We propose a simple state observed based implementation of this approach. Simulation results, where a HRP-2 robot tries to imitate another HRP-2 robot, show that imitation can be realized with virtually no delay, in perfect synchrony.