Design and implementation of a self-driving car using deep reinforcement learning: A comprehensive study

Abstract

This paper presents a groundbreaking and comprehensive study on the design, implementation, and evaluation of a self-driving car utilizing deep reinforcement learning, showcasing significant advancements in autonomous vehicle technology. Our robust framework integrates three innovative AI models for essential functionalities: road detection, traffic sign recognition, and obstacle avoidance. The system architecture, structured around a three layers “DDD” (Data, Detection, Decision) approach, involves meticulous data preprocessing for traffic signs and road data, followed by specialized Deep Learning models for each detection task, including a CNN for traffic signs, a CNN for road detection, and the pre-trained MobileNet-SSD for obstacle detection. A reinforcement learning agent in the Decision Layer processes these outputs for real-time control (steering, acceleration, braking) through a continuous learning process with environmental feedback. The research encompasses both extensive simulation in Unity, leveraging the ML-Agents toolkit for agent training across diverse environments, and crucial real-world deployment. Our reward/punishment system in the simulation environment, based on collisions with road markers and obstacles, refined the agent’s decision-making. The trained AI models were successfully exported and deployed onto a physical prototype, controlled by a Raspberry Pi and equipped with a camera and ultrasonic sensors. Real-world testing affirmed the robust performance of the physical model in detecting roads, recognizing traffic signs, and effectively avoiding obstacles. Quantitative results demonstrate compelling performance, including over 90% accuracy in obstacle detection and a 15% improvement in navigation efficiency compared to traditional algorithms under controlled simulation conditions. Model evaluation metrics show a 98% accuracy, 12% loss, and a prediction rate exceeding 77%. This study not only contributes a comprehensive framework for autonomous vehicle development but also highlights the transformative potential of deep reinforcement learning for creating intelligent and adaptable autonomous systems in both virtual and real-world scenarios, paving the way for safer and more efficient transportation technologies.