Research On Obstacle Avoidance Of Mobile Robots In Multi-dynamic Obstacle Environment Based On Deep Reinforcement Learning

Mobile robots has improved the production efficiency of the society,The difficulty of mobile robot is how to avoid collision.In particular,in a dynamic obstacles environment,there is no communication between mobile robot and dynamic obstacles,the goal of dynamic obstacles cannot be known by mobile robot.In addition,in order to find the shortest path of mobile robot to its goal,it is usually necessary to consider the interaction with its dynamic obstacles.there are many methods to solove the collision problem.such as the analytical method optimal reciprocal collision avoidance.However,these methods have many disadvantages,such as large computation,poor real-time performance,and poor stability when increasing the number of dynamic obstacles.In order to overcome these shortages,this paper proposed a method based on deep reinforcement learning to avoid collision between mobile robot and dynamic obstacles.This paper mainly studies the collision avoidance of mobile robot.I designed deep reinforcement learning method based on the characteristics of the mobile robot and the tasks to be completed,some observable information is used to coordinate with each other to reach its goal without communication between mobile robot and dynamic obstacles.Based on the Deep Q-network,This paper designs a reinforcement learning model for collision avoidance.The the structure of DQN is improved by the Long Short Term Memory neural network is introduced to preprocess the state,and by setting of the experience pool is modified.The effiency of the improvement is validated through simulations,in the simulation results,the time for mobile robot to reach its goal becomes shorter.In order to keep the algorithm stable when the number of dynamic obstacles increases,an attention mechanism was introduced to make the mobile robot pay more attention to the more influential neighboring dynamic obstacles.The results showed that by using the algorithm including attentional mechanism,mobile robot could still reach the target point without collision in a short time when increasing the number of dynamic obstacles.This algorithm was more more stability than the traditional method and the deep reinforcement learning method based on Long Short Term Memory neural network.