Research On Path Planning For Mobile Robots Based On Deep Reinforcement Learning

Mobile robots can replace humans in complex and repetitive operations,avoiding human physical work to a large extent.The application of mobile robots is valued in the industrial upgrading of various industries and fields.The most important basic function in mobile robotics is the realization of path planning for mobile robots.Therefore,it is important to design autonomous path planning strategies for robots that can cope with static and dynamic environments while improving the efficiency of robot path planning.Based on this,the specific research in this paper is as follows.Firstly,to solve the problem of "dimension explosion" when reinforcement learning methods are used in path planning in complex environment,a path planning algorithm based on deep reinforcement learning is proposed.By establishing end-to-end learning methods,deep Q network is used to fit Q values.At the same time,in view of the overestimation problem caused by the deep reinforcement learning algorithm in path planning,it is proposed to use the action selection strategy of optimizing the target to improve the radical operation of max when selecting the target.Secondly,in order to store more excellent samples in the experience pool and improve the learning ability of mobile robots,the method of PER sampling during experience playback is proposed to improve the retention number of excellent experience samples in the experience pool by setting the baseline and bias value,and then the experience samples are sorted according to the importance,so as to ensure that the experience samples containing excellent information can be used more frequently.Then,in order to further improve the efficiency of path planning training for mobile robots,transfer learning strategy is introduced into the improved algorithm to reduce the requirements and difficulties of training data collection.Finally,in order to make the research content more consistent with the realistic environment,a 3D perspective simulation platform is built based on Gazebo,which is integrated with the mobile robot model designed based on ROS,and a variety of obstacle experimental simulation environment was designed.Through experimental simulation and comparison of experimental data in the training process,it can be concluded that the algorithm improvement points proposed in this paper obtain higher average reward value and better convergence when applied to the path planning of mobile robots.Moreover,the path planning length is shortened and the training efficiency of mobile robots is improved by comparing the path trajectory with the data in the path planning task.So that it can more smoothly plan the collision free movement path.