This paper presents a novel interactive motion planning system for assembly/disassembly operations. Our system consists of three layers: interaction layer, learning layer and motion library layer. In interaction layer, user’s manipulation in difficult scenario is liberated by relaxing collision constraints. The resulting path is retracted and connected by random retraction method and BiRRT algorithm. A motion path which successfully passed through the narrow passage or information of geometrical interference in failed case is returned to user. In learning layer, motion primitives corresponding to prior similar scenario are selected by scenario comparison which is based on medial axis, and then transformed to generate new motions. Significant improvement for motion planning of non-convex object in challenging scenarios with narrow passages is obtained by interactive process. The introduction of learning mechanism can reduce global planning time and obtain experiential knowledge.