Deep Reinforcement Learning Based Mobile Robot Navigation Using Sensor Fusion

Abstract

At present, mobile robot navigation usually uses the traditional SLAM method, which can accurately obtain the environment information, but it is difficult to effectively model the environment if the environment changes greatly in the later stage or in dynamic and complex scenes. In order to make the mobile robot better adaptable to the environment, a reinforcement learning method is used to control the mobile robot. This paper mainly introduces the application of reinforcement learning algorithms and data fusion to mobile robot navigation obstacle avoidance. We use the deep deterministic policy gradient(DDPG) algorithm in reinforcement learning to train neural networks. Gazebo simulation platform is adopted as the simulation environment of this paper, which can effectively simulate indoor and dynamic pedestrian environments. Considering that the depth value detected by a binocular camera may be affected by light in a real environment, we supposed the method of sensor fusion which is added to fuse the data of the laser sensor and vision sensor for 3D reconstruction and generate corresponding point cloud data. It is verified that this method performs better than the vision sensor as well as the laser sensor, which can improve the success rate of navigation by 10% in a dynamic environment. And this method can make the trajectory of the agent smoother and reach the target point faster.