It is important for collaborative robots to share goals with a human partner, but considering all the possible goals in advance is difficult. Therefore, collaborative robots are required to generate correct action towards new unlearned goals. In this research, we focus on a method of generating an action by providing a goal image. In order to acquire the relationship among the goal images, the corresponding visual information and the robot’s joint angles, Long Short-Term Memory Recurrent Neural Network was used. We confirmed that a robot with our proposed method was able to generate correct actions even when unlearned goal images were given.