— We examine the usefulness of passive compliance in a manipulator that learns contact motion. Based on the notice that humans outperforms robots with the contact motion, we follow two aspects of human manipulation: passive compliance and learning. As imitation of human’s arm and learning, we use a robot arm with passive compliant joints and it learns a policy for peg in hole by the proposed gradient descending method. We present that the passive compliance provides with quick and stable learning as well as a slow control sampling time.