Self-Driven Vehicle: Learning to Drive in a Simulated Virtual Environment

As autonomous vehicles move closer to real-world deployment, the ability to train them safely and effectively in a wide range of traffic conditions has become increasingly important. Simulation platforms have emerged as a powerful tool for developing and testing self-driving systems, offering safe and cost-effective alternatives to real-world trials. However, most existing simulators are built around structured road layouts and rule-based traffic behavior typical of Western countries, failing to represent the chaotic and unstructured driving environments often seen in developing regions. This paper presents a high-fidelity simulation framework built using Unreal Engine 5 to better reflect these real-world complexities such as faded or missing lane markings, jaywalking pedestrians, mixed traffic types, and highly unpredictable driver behavior. Within this environment, an autonomous vehicle agent is trained using deep reinforcement learning. The agent learns to operate effectively by interacting with diverse scenarios, developing robust policies for steering, acceleration, braking, and high-level decision-making in unpredictable contexts. Experimental results show that the agent successfully navigates long episodes without collisions and adheres to traffic norms despite the simulated chaos. The study demonstrates the value of region-specific, unstructured simulation environments in preparing autonomous systems for globally diverse and challenging real-world conditions.