Research On Dynamic Obstacle Avoidance Algorithm Based On Reinforcement Learning

With the continuous development of technology,more and more robots are being employed in social production,particularly mobile robots that often work in dynamically complex environments such as airports and shopping malls.many of the previous dynamic obstacle avoidance algorithms have typically neglected the future behavioral states of dynamic obstacles.Instead,they have relied on simplistic geometric structures or sampling methods to generate avoidance strategies.In dynamic environments,these traditional methods often result in unnatural collision behaviors by the robot due to the randomness of obstacle movement,leading to frequent changes in control signals and making it difficult to meet the practical requirements of the application.Therefore,in order to meet the needs of school-enterprise cooperation companies for mobile robots to efficiently perform navigation tasks in dynamic and complex environments,this thesis proposes a deep reinforcement learning-based obstacle avoidance algorithm that combines obstacle behavior prediction with robot motion planning.Through direct acquisition of the robot’s control output signal based on environmental states,collision between mobile robots and obstacles can be avoided,while ensuring that the robot can approach the target point as quickly as possible.This thesis addresses the existing problems in the research of mobile robot obstacle avoidance and conducts in-depth research on the application of deep reinforcement learning in complex and unknown obstacle avoidance scenarios from multiple perspectives.The specific research contents are as follows:First,in response to the problem of mobile robots being prone to collisions in complex environments,this thesis proposes an improved DDPG obstacle avoidance algorithm model.The algorithm model introduces an LSTM network on the policy network of the DDPG algorithm,extracts key features from laser radar data,and forgets unimportant information,enabling the mobile robot to focus more on dynamic obstacles that have a greater impact on it,especially those that are nearby,in order to make decisions and enhance the accuracy of obstacle avoidance decisions.The improved algorithm model has been tested in multiple complex simulation environments,with an average obstacle avoidance success rate of 75% in comprehensive performance.Secondly,to address the issues of poor algorithm stability and weak generalization ability,this thesis proposes an obstacle interaction model that combines obstacle behavior prediction with robot motion planning.The model extracts useful obstacle interaction feature information for obstacle avoidance by encoding obstacle interaction information,and extracts time series features of obstacles to reduce the short-sightedness of obstacle avoidance strategies.This enhances the neural network’s analysis and understanding of input feature information,allowing it to learn potential rules between obstacles,understand and predict the motion trajectories and states of obstacles,and ultimately make more accurate decisions and controls,thereby enhancing the stability and generalization ability of the algorithm.Third,this thesis presents an improved obstacle avoidance algorithm based on a local attention mechanism.To address the problem of deep neural networks not being able to fully utilize important information in the environment,a local attention mechanism is introduced on the neural network of DDPG to allow the neural network to focus more on important information and improve its understanding of the environment.Moreover,this attention calculation mechanism can reduce redundant computation,thereby improving the computational efficiency of the network.Finally,through simulation experiments and comparison with traditional algorithms,the improved DDPG-OIM-LAM algorithm was found to have an average obstacle avoidance success rate of up to 95% in various complex obstacle avoidance scenarios,demonstrating higher stability and generalization performance.