We constructed virtual agent system for acquisition of robot's behavior. Moreover we focus on Observation Learning. Observation learning was based on gestual interface. Recently, robots that fit into the human community are needed. These robots need natural motions, which are understandable by users for natural interaction with human. For robots acquire natural motion, we used observation learning. A robot acquires a natural motion, which adapted a user's taste from gestural interaction between a user and a robot. However a robot should move many times until acquired behavior which a user feels satisfaction. Observation learning using real-robot was burdened robot body. Acquired behavior using a robot was not able to reused for other robots. So we constructed virtual agent basis for several robots. A virtual agent acquired important points in a motion. A real robot moves using important point, which was acquired the virtual agent.