Research On Path Planning Method Of Indoor Service Robot Based On Deep Reinforcement Learning

With the prosperous development of robotics in recent years,service robots have become more and more inseparable to human life.Service robots often work in complex and unstructured environments,so they need high-precision path planning abilities.The advantages of traditional path planning algorithms have clear logic and simple implementation,but with the increasingly complex structure of the environment faced by the robots,the disadvantages of traditional algorithms,such as poor adaptability to environmental changes,low replanning efficiency and inability of self-adjusting parameter,all make them unable to meet the path planning requirements of indoor robots.In order to enable indoor service robots to effectively cope with unstructured environments,this thesis investigates the problem of path planning based on deep reinforcement learning method.The main works of this thesis are:(1)To address the problem that indoor service robots using Dynamic Window Approach(DWA)unable to self-adjust the weight factor due to environmental changes,and easy to fall into the local optimal solution,a Proximal Policy Optimization-Dynamic Window Approach(PPO-DWA)method is proposed by this thesis.A novel deep neural network is designed to obtain the relationship between environmental information and weight increments in the DWA evaluation function.An unstructured random obstacle training environment is constructed,and the network is trained by proximal policy optimization method to make the algorithm adaptive to environmental changes.Through the optimal weight increment output by the neural network,the robot has the ability to adapt to the environment.(2)To address the problem that the indoor service robot uses traditional algorithm for global path planning,the robot is prone to collision with obstacles during the navigation due to too small obstacle clearance,this thesis applies the Dijkstra search method based on Voronoi diagram.The environment is modeled by Voronoi diagram to reduce the number of path search nodes and ensure the obstacle clearance of each node is large enough.The Dijkstra method is used to traverse the nodes in the diagram to ensure the connectivity between the start point and the target point,and get a global optimal path.(3)To verify the performance of the PPO-DWA proposed by this thesis,a simulated service robot for navigation and a virtual restaurant for comparison experiments are built on the Gazebo platform.Through simulation experiments,the results show that the PPO-DWA method proposed in this thesis can significantly improve the local planning performance of indoor service robots in unstructured environments,and the planned motion trajectories are smooth and collision-free.