In this paper we survey a number of methods for teaching human motion/force skills to robots. Since verbal description of a human skill could only provide qualitative guidelines for machine programming, several works were motivated to identify human skills analytically through the study of human teaching data. The key advantage of direct teaching by human demonstration is that the method is highly user-friendly; it requires no special knowledge of machine programming from the users, and it automatically extracts user intentions and strategies from task performance demonstrated by the users. From the early teaching/playback method to the recent methods of teaching sensory feedback laws, a historical perspective of the field of human skill teaching is given in this paper. Limitations and applicability of methods under review are discussed. Essential issues such as representation and identification of human skills as well as consistency of teaching information are addressed.