AutoNav: A CARLA-Based Framework for Autonomous Vehicle Simulation

Abstract

The paper explores the way that the design of autonomous vehicle simulator creates a virtual environment inside which the high-cost and safety-innovative, autonomous vehicle (AV) technologies could be achieved through innovation and experimentation. CARLA (Car Learning to Act) belongs to the most active autonomous driving simulators that are open-source. It is created on Unreal Engine and high 3D environments will recreate reality of urban and highway backgrounds. CARLA is simulated that it has bases LiDAR, RGB cameras, radar, GPS and IMU sensors mimicking a sensor perception system found in the visual representation of a real-world AV. It also accommodates areas of dynamism, namely cars, pedestrians, traffic lights, and changing weather, which imply that it is best to test AV systems under various conditions. CARLA provides the ability to have controlled experiments to increase the safety and reliability of AVs. Experimental simulations in AutoNav achieved consistent lane tracking accuracy above 95% and reduced collision occurrences by 80% compared to baseline models. The system also demonstrated smooth obstacle avoidance and efficient route completion under varying weather and traffic conditions.