Research On Active SLAM Algorithm Based On Deep Reinforcement Learning In Complex Environment

In order to make the robot work better in the complex indoor environment,it is necessary for the robot to have the ability of autonomous navigation,obstacle avoidance and map construction.Traditional SLAM(Simultaneous Localization and Mapping,SLAM)algorithm for planning needs to manually set the moving trajectory of the robot,which increases labor costs and makes the construction of the map more complicated..The initiative to SLAM(Active SLAM)algorithm based on the understanding of the environment,so that the robot can realize independent mapping,instead of the needs of mapping of the traditional manipulation robots,which improves the initiative and autonomy of the robot.In view of the above problems,this paper studies the obstacle recognition method based on depth image of mobile robot,obstacle avoidance based on depth reinforcement learning algorithm and active SLAM combining obstacle avoidance with FastSLAM map construction.First of all,the deep learning network FCRN(Fully Convolutional Residual Networks)algorithm is used to identify obstacles to get depth image and build experimental environment of a two-dimensional map.The harris-lebot2 Gazebo simulation model is set up,and the RGB image is predicted with FCRN depth based on Kinect,so that the robot can obtain the obstacle and environment information.As can be seen from the loss function of training and testing,with the increase of the number of steps,FCRN depth image prediction effect is better,and the environment map is constructed by FastSLAM method.Secondly,the obstacle avoidance algorithm is integrated into the SLAM framework,and an active SLAM method based on deep reinforcement learning is proposed.Based on the FCRN algorithm to identify the depth image obtained by obstacle training,the Dueling DQN algorithm is used to plan the obstacle avoidance path,and merged into FastSLAM to complete online obstacle avoidance and map construction.Experiments show that in the environment with different number of static and dynamic obstacles,the proposed method can effectively avoid obstacles in different states in the process of map construction and realize the autonomous navigation of the robot in the complex environment.Finally,in order to improve the speed of depth image prediction and optimize the robot's walking path,an improved active SLAM method based on Monodepth image recognition is proposed.Monodepth algorithm is used to improve the prediction speed of obstacle recognition in depth image,and the obtained binocular stereo image replaces the explicit depth data used in FCRN algorithm training.Dueling DQN deep reinforcement learning based on probability selection is used to optimize the walking path.Finally,the fusion with FastSLAM completes the autonomous navigation under different number of static and dynamic obstacles and constructs the environment map.