Automatic vehicle parking using an evolution-obtained neural controller

Abstract

Within the problems that can be solved with autonomous robots, automatic parking is an area of great interest, since it presents a complex scenario where the agent must go through a series of obstacles to reach its goal. Existing solutions usually require some kind of external mark for monitoring or global vision that indicates where the agent is at a given time. This article presents an evolutionary strategy to generate a robotic controller based on a neural network that successfully solves the problem of vehicle parallel parking using only local information. The performance of the tness function is analyzed, focusing not only on the agent reaching its goal, but also on it doing so in a manner that is appropriate for the physics of a vehicle. Additionally, the Player/Stage simulator is broadly discussed, since it is one of the most widely used simulators nowadays in robotics.