This study focusing on peg-in-hole task, which is the basic elements of assembly. We propose a method for combining human knowledge with machine learning to generate robot motions. This method aims to acquire a robot motion with a high success rate in a short learning time under the condition of position and orientation errors.