Deep Reinforcement Learning for Automated Car Parking

Description

This article explores the development of a Deep Reinforcement Learning (DRL) -based agent able to perform both path planning and trajectory execution, processing sensor perception information and directly controlling the steering wheel and the acceleration, like a normal driver. As a preliminary investigation, we limit our research to low-speed manoeuvers, in a challenging narrow drivable area. The vehicle's agent completely relies on the real-time information from the sensors, thus avoiding the need of a map. We show the validity of the proposed system in a simulated car parking test, in which the agent has been able to achieve high target reach rates, with a limited number of manoeuvers (gear inversion rate), outperforming the well-established Hybrid A-Star path planning algorithm in both the metrics. Further research is needed for improving the generalization ability of the agent and its application in more dynamic driving environments.

Additional details