End-to-End Deep Learning Approach for Comprehensive Autonomous Driving in Simulated Environments

Developing an end-to-end deep learning-based navigation system for autonomous vehicles simplifies the traditional modular pipeline by directly mapping sensory inputs to control outputs. In this study, a neural network processes RGB images from a front-facing camera and distance measurements to generate steering, throttle, and brake commands. The system was trained and evaluated in the CARLA simulator, specifically in Town 04, under diverse weather conditions (e.g., clear sunset, heavy rain) and scenarios involving static and dynamic obstacles. Using convolutional and dense layers, the proposed architecture achieved an autonomy score 100%, demonstrating robust navigation, precise obstacle avoidance, and adaptability to challenging scenarios. These results underscore the potential of end-to-end approaches to enhance the efficiency and scalability of autonomous driving systems, providing a solid foundation for future real-world implementations.