Autonomous Mobile Robot with Simple Navigation System Based on Deep Reinforcement Learning and a Monocular Camera

The purpose of this study is development of an autonomous mobile robot navigation system based on deep reinforcement learning with a monocular camera, without 2D-LiDAR. The proposed system is based on DDQN(Double Deep Q-Network) as deep reinforcement learning. The system requires the input data as states of DDQN that include the range data around the robot. In this paper, the range data is estimated from a monocular camera instead of 2D-LiDAR. Monocular camera is relatively cheap compared to LiDAR, which can lower the hurdles for spreading robots in the world. The proposed system converts the depth images estimated from monocular camera to 2D range data that is input to the learned model based on 2D plane. The learning on 2D plane is effective to obtain stable models from deep reinforcement learning. Then, we conduct two experiments and evaluate the proposed system. The results show the autonomous navigation was achieved according to camera image-based states.