A reinforcement learning-based reverse-parking system for autonomous vehicles

This work presents the design and implementation of a reinforcement learning-based autonomous parking system where an agent is trained to reverse-park in a selected parking spot. The parking procedure is divided into three stages, and each stage has its corresponding surrogate objective that contributes to the overall parking process. The model solely depends on features extracted from a top-view image of the parking space. It has the advantage of potential deployment in smart parking buildings without refitting non-autonomous cars with modern sensors. The training was conducted offline on a simulation utilizing the proximal policy optimization algorithm. The model was then transferred and tested on a hardware prototype of the parking space. The results of the system were successful as the successful parking rate reached 100% with no collisions with any objects, and the fastest parking time reached 10 s. The testing was conducted on multiple samples and scenarios of the parking setup.