Research On Parking Trajectory Planning Based On Reinforcement Learning

With the improvement of vehicle intelligence and networking,automatic parking technology has received more and more attention.At present,due to the poor parking environment and insufficient driver experience,it is difficult to park in a small space.Therefore,how to achieve parking in a small space is of great significance.This paper studies the path planning in the automatic parking system,and proposes to use reinforcement learning algorithm to explore the parking path.In the vertical working condition,by designing three-stage trajectory rewards,the training can obtain the vertical parking trajectory;in the horizontal working condition,by designing the two-stage trajectory rewards,the training can obtain the horizontal parking trajectory.A real vehicle test was carried out to verify the feasibility of the parking trajectory.First,in order to apply reinforcement learning for parking trajectory planning,according to the parking motion characteristics of the vehicle,the reinforcement learning state transition equation is constructed based on the vehicle kinematics model;possible collisions are analyzed,collision safety judgment rules are established,and parking The end of the car is designed to reward the goal.Through simulation,it is found that the single-objective reward design will guide the vehicle to approach the goal first and then avoid obstacles,resulting in a long parking path.Therefore,a piecewise reward function design method is proposed.Secondly,for vertical parking trajectory planning,according to the driver’s parking process,the vertical parking process is divided into three stages to design trajectory rewards.From the perspective of obstacle avoidance,the applicable scope of each trajectory reward is defined and simulated.Train and compare the training results with the multi-arc parking trajectory.Further,in order to solve the problem of large parking space,the penalty value of the frequency of the control amount change is reduced to reduce the length of the forward trajectory,and the collision penalty value is reduced to increase the exploration space.Simulation training is performed to obtain the parking trajectory,and the obtained Q is verified.The matrix can reduce the parking space and complete parking in different sizes of parking spaces.In horizontal parking planning,the reward function is divided into two segments according to the change of heading angle during parking.Aiming at the problem of large parking space,according to the constraints of vehicles and parking spaces,the scope of application of two trajectory rewards is defined.The parking trajectory is obtained through simulation training,and the verified Q matrix can shorten the parking space and complete parking in different sizes of parking spaces.In order to verify the effectiveness of the actual vehicle trajectory obtained by training,according to the performance requirements of the parking system,a number of different initial vehicle orientations were designed,and the actual vehicle test verification was carried out under different parking spaces.The test results show that the learned parking trajectory is effective under vertical and horizontal working conditions;the vehicle can park smoothly and without collision under the condition that the width of the parking space is met.