Since it is quite difficult to create motions for humanoid robots having a fairly large number of degrees of freedom, it would be very convenient indeed if robots could observe and imitate what they want to create. To this end, this paper discusses how humanoid robots can learn through imitation taking into consideration the fact that demonstrator and imitator robots may have different kinematics and dynamics. As part of a wider interest in humanoid motion generation in general, this work mainly investigates how imitator robots adapt a reference locomotion gait copied from a demonstrator robot. Specifically, the self-adjusting adaptor is proposed, where the perceived locomotion pattern is modified to keep the direction of the lower leg contacting the ground identical between the demonstrator and the imitator, and to sustain dynamic stability by controlling the position of the center of mass. The validity of the proposed scheme is verified through simulations on OpenHRP and real experiments.