Reinforcement learning algorithms in highway driving

Abstract

This thesis examines the potential of employing photorealistic synthetic data for training neural networks to achieve collision-free autonomous driving in highway environments. Centered around a case study of a highway accident involving a Tesla Model 3, the research utilizes Unreal Engine 5 to reconstruct the event and explore the effectiveness of an End-to-End autopilot system. The study innovatively applies Domain Structured Randomization to generate varied driving scenarios, assessing the autopilot’s adaptability and response. The absence of real-world domain access underscores the significance of synthetic data in simulating and analyzing the incident, aiming to enhance the safety features of autonomous driving technologies.