In this paper, we proposed learning-based motion planner for autonomous mobile robot which outputs discrete motion commands by referring from the onboard RGB image and target position without using map. It is difficult to get policy for navigation of mobile robots using only visual sensor. In recent years, research on visual navigation using deep neural networks has been conducted. However, these methods do not have a structure to store time-series information in a network. Therefore, this paper presents the DNN motion planner with LSTM trained by using expert trajectories. As a result, we show that our method can navigate agent in simulation environment without map.