A Data-driven Comparative Study of End-to-End and Rule-based Control for Autonomous Driving

This paper presents a data-driven comparative study of two control approaches in autonomous driving: a traditional rule-based system and a ResNet-18-based end-to-end deep learning model. The rule-based system utilizes YOLOv8 segmentation to detect lane markings and generate steering and speed commands through handcrafted control rules. Driving data - including camera images, steering angles, and motor speed commands - was collected from this system in real-time under various driving conditions.Using this dataset, we trained a ResNet-18 regression model to predict both steering and speed directly from visual input. The end-to-end model was evaluated in both simulated and real-world environments. Results show that it effectively replicates rule-based behavior, with high prediction accuracy in steering and acceptable performance in speed control. However, without additional fine-tuning, the model exhibited instability in real-world scenarios, especially when deviating from ideal trajectories. Fine-tuning with recovery data significantly improved performance, enabling stable lane-following behavior using model-inferred commands.These findings highlight the importance of data diversity and quality in end-to-end learning and demonstrate that deep learning models, when properly trained, can achieve control performance comparable to rule-based systems.