Research On Autonomous Valet Parking Path Planning Method Based On Deep Reinforcement Learning

Autonomous valet parking has the characteristics of closed parking lot application scenes and low speed of vehicles in the scenes.It is considered to be the first technology to be applied in the field of autonomous driving.In the autonomous valet parking system,path planning is a very important component,the main purpose is mainly used to calculate an obstacle avoidance path from the starting point outside the parking lot to the target parking space.With the increase of car parc,the uncertain behavior of vehicles increases,and the parking environment becomes more and more complex.At this time,traditional path planning methods have problems such as slow response speed and low planning efficiency,and it is difficult to cope with dynamic changes,which is intensified.Learning directly interacts with the environment,has strong self-learning ability,and can complete continuous planning tasks in complex environments.Therefore,it is of practical significance to study the application of deep reinforcement learning in autonomous valet parking.Based on the above background,this article proposes a TD3 path planning algorithm based on deep reinforcement learning for the task of autonomous valet parking path planning,and tests it in a parking lot simulation environment.The experimental results show that the TD3 algorithm is more convergent than other deep reinforcement learning algorithms.The speed is faster,but it is limited to the smallscale parking lot simulation environment;in order to realize the path planning task in the larger-scale simulation environment,this paper proposes the Leader＿TD3hierarchical deep reinforcement learning path planning algorithm,which is based on the parking lot prior map,the upper layer Using the method of "Connected region analysis + Expansion layer + PRM",the global route is divided into local paths composed of sampling points,and the sampling points are fed back to the underlying TD3 deep reinforcement learning algorithm as a leader,and the corresponding is completed by the well-trained TD3 algorithm Local path planning between adjacent sampling points.In order to verify the feasibility of the algorithm,this paper builds a 3D parking lot simulation environment with a physics engine based on Gazebo.In view of the long training period and difficulty of convergence in the 3D simulation environment of the algorithm,this paper builds a 2D parking lot simulation environment with the observation state and the reward function consistent,and arranges the initial development and optimization of the algorithm in the 2D simulation environment.Use transfer learning to transfer the best algorithm model trained in the 2D simulation environment to the 3D simulation environment.The experimental results prove that this transfer learning method significantly improves the convergence of the algorithm.In order to verify the superiority of the algorithm,this paper selects the "A*+DWA" traditional path planning algorithm to compare with the algorithm proposed in this paper,and develops evaluation indicators for the task of this paper.The experiment is carried out in different scale simulation environment.The experimental results show that the Leader＿TD3 hierarchical deep reinforcement learning path planning algorithm has certain advantages in completion rate,total planning length and planning time compared with the traditional algorithm,and can deal with the complex parking environment.