Research On Reinforcement Learning Algorithm For Mobile Vehicle Path Planning In A Special Traffic Environment

With the continuous progress of intelligent information processing technology,automatic driving has become a research hotspot for experts in recent years,and path planning is the key technology.Intelligent path plarning can greatly reduce the completion time of mobile tasks and save labor and material costs,which has great practical value.In this paper,based on the actual project requirements,PyQt5 framework is used to build a simulation platform to meet the characteristics of the scene for the special traffic environment studied.This paper mainly designs two kinds of path planning algorithm based on reinforcement learning.One is to use DDQN algorithm in deep reinforcement learning to train the moving path by designing the standard moving motion of vehicles.The other is to propose an improved three-stage path generation algorithm and path evaluation system on the basis of using third-order Bessel curve to simulate the motion trajectory,and then train the variable parameters in the three-stage algorithm with DDPG algorithm in deep reinforcement learning to realize the optimization of the moving path.The main work of this paper includes the following aspects:(1)Scene modeling.This paper uses PyQt5 framework to design and implement scene modeling tool in PyCharm.The simulation environment is built by imitating the actual scenes of flight deck and storage warehouse.The main features of the environment are boundary contours,dockable positions,obstacles and moving objects.(2)Design and implement the path generation algorithm under the standard action mode.In this paper,several angles are selected at equal intervals within the feasible Angle range of the vehicle,and the arc motion is made at a certain turning radius and Angle at each Angle to form a discrete action set.Set the reward function and use DDQN algorithm for training.(3)Design a three-stage path generation algorithm based on Bessel curve.The algorithm consists of two parts:generation and evaluation.The generation part adopts the three-stage method of "arc-shaped exit station location-Bessel curve transfer-arc-shaped entry station location".The evaluation part uses entropy method to determine the weight coefficient of each evaluation index,and takes the distance method of advantages and disadvantages as the final evaluation function.(4)The training optimization of the three-stage path generation algorithm.In the three-stage generation algorithm designed in this paper,there are several variable parameters that determine the final generation path.Based on this,markov decision process model is constructed and DDPG algorithm is used for training to determine a set of optimal parameter values and optimize the generation path under the three-stage algorithm.In this paper,the scene modeling tool is used to build a number of scenes for algorithm research,and a lot of simulation experiments are carried out for the two algorithms.The experimental results show that both algorithms are effective and can plan an ideal path for any given moving task in the scene.