Self-Learning Parking Assistant Using Reinforcement Learning in a Mini-Car Simulator

This paper presents a self-learning parking assistant which is created and tested in a mini-car simulation environment. The system uses a Deep Reinforcement Learning (DRL) agent to autonomously learn how to do complex left/right parallel and perpendicular parking. The DRL agent uses Proximal Policy Optimization (PPO) to produce steering and velocity commands using virtual ultrasonic range data and proprioceptive vehicle state. A reward function has been designed to penalize the DRL agent for colliding with any object, excessive maneuvering, and having a large final pose error; conversely, it will reward the DRL agent for completing the parking successfully. This study demonstrates that the proposed self-learning parking system is successful at approximately 98.5% of the time when parking in randomized environments with different types of obstructions. Additionally, the DRL agent demonstrates competitive performance results in terms of the time it takes to perform a maneuver (average) over the course of the experiments, which indicates that using (DRL) for autonomous real-time parking within limited spaces is a feasible approach. The development of new and emerging concepts such as federated learning and Proof of Stake (PoS) blockchain has also been integrated in order to provide secure sharing of data, and validate the model of the proposed self-learning parking system. Finally, the completed design, training pipeline, evaluation of experimental results, and future direction of the self-learning parking system will be discussed in this paper.