Research On Path Planning Method Based On Deep Reinforcement Learning In Indoor Environment

With the gradual development of automobile automatic obstacle avoidance technology and service robots,path planning has become a hot issue in the field of mobile robot technology.Deep reinforcement learning in the field of artificial intelligence does not need artificial markers and does not rely on prior knowledge,so many fields have been studied and developed in combination with practical applications,and there are also researches in this direction in robot path planning tasks.In this paper,monocular camera is used as the sensing method of robot to study the indoor obstacle avoidance problem based on deep reinforcement learning.Firstly,on the basis of the encoder-decoder network structure,using the method of supervised training,a method that can use monocular vision to achieve high-throughput lightweight depth estimation on embedded systems is proposed.Jetson TX2 is the target embedded platform.Effective validation of the model and comparative analysis with other Depth estimation algorithms are completed on NYU Depth V2 of the exposed data set,indicating the superiority of the algorithm.Secondly,on the basis of depth of reinforcement learning theory has carried on the thorough research,aiming at DQN algorithm is easy to make the Q value estimates,the characteristics of slow convergence,and considering influence factors in the original DQN algorithm one-sided and ignored the environment status of value,the original DQN algorithm was improved,using the improved algorithm,compared with the original DQN network experiment shows the advantage of the method.Then,the ROS system and Gazebo simulation software were used to build an experimental environment for path planning simulation,which was used for training the static obstacle avoidance algorithm.Then the Turtlebot2 model of the simulation robot was added to the simulation environment,so as to complete the construction of the indoor robot path planning simulation platform.Finally,the Turtlebot2 mobile robot was trained in the deep reinforcement learning algorithm model in the established comprehensive simulation platform,and then tested in the real environment,which verified the feasibility of the algorithm in this paper for path planning task in indoor environment.