A method for representing human skills in assembly tasks so that they can be transferred to intelligent robotic systems is proposed. Data from which control strategies of the human operator are extracted are acquired by a direct teaching method. Two levels of control strategies are extracted from the data: one is a high-level sequence of motion with regard to discrete state transitions of the contact state, and the other is compliant motion control strategies, with which the human operator manipulates parts. Based on a priori knowledge of the geometry of parts, the authors first generate a state network representing possible transitions of contact states. The human control data are interpreted with reference to this network and a motion sequence is extracted from the data. The compliant motion control strategy necessary for each state transition is then identified by processing the force and velocity profiles of the acquired data. The identified control strategies in individual steps of the operation are then compiled in order to simplify the representation of the strategies as well as to obtain smooth state transitions. Experimental results illustrate the method.