In this study we focus to develop novel approach to autonomous driving research, emphasizing multi-task learning (MTL) with a single input image. The proposed end-to-end deep learning architecture incorporates shared layers to efficiently handle statistically correlated tasks, including estimating steering angle, braking, acceleration, and the number of lanes on both sides of the vehicle. To address the challenge of varying task significance and ranges, the model explores different normalization schemes, identifying the inverse validation-loss weighted scheme as the most effective. The evaluation conducted on four publicly available datasets and a synthetic dataset (GTA-V) showcases the superior performance of the proposed approach compared to current state-of-the-art methods. This highlights the potential of the end-to-end deep learning architecture and multi-task learning for advancing autonomous driving research.