End-to-end learning offers a compact alternative to modular autonomous driving pipelines by directly mapping sensory observations to control actions. This paper presents a behavioral cloning framework that learns continuous steering control from front-facing monocular camera images using a deep convolutional neural network trained on 1,230 simulation driving records. The task is formulated as supervised regression over steering angles, with preprocessing and augmentation designed to address a 76.6 % steering imbalance toward straight driving. The system integrates image normalization, geometric steering correction, and photometric augmentation, and is deployed in a real-time simulator–client architecture with speed-regulated throttle control. Closed-loop evaluation demonstrates stable autonomous lane following and curve negotiation, with augmentation significantly improving recovery behavior and validation loss stability. These results highlight the importance of data-centric design and distribution-aware training in end-to-end control systems. The proposed framework provides a reproducible foundation for extending behavioral cloning toward temporal modeling and sim-to-real transfer.
Kirubakaran et al. (Fri,) studied this question.