The reinforcement learning application for the automatic parking of car in a simulated environment

Abstract

In the present work, a reinforcement learning solution is proposed to carry out automatic perpendicular parking  in a 4-wheeled vehicle. It focuses on designing a reward function used to train the Proximal Policy Optimization and Soft Actor-Critic algorithms and combine both in an assembly. Training the algorithms with random starting positions gives poor performance. Finally, it is possible to obtain a success close to 99 % and an absolute deviation of approximately 1 degree.