In this paper, we propose a kind of "teaching by demonstration" method, aiming at its application to humanoid robots at home in the future. The demonstrator's motion is captured by a pair of stereo cameras mounted on his/her head, locating very close to his/her eyes. By tracking the landmarks attached to the demonstrator's hand and the working environment, one can estimate not only the demonstrator's hand motion but also his/her head motion, which can be used for the active vision system. Experimental result shows the effectiveness of the proposed framework.