A Transfer-Learning-based Strategy for Autonomous Driving: Leveraging Driver Experience for Vision-Based Control

This paper explores the utilization of a novel transformer-based architecture for end-to-end learning in predicting steering angles in self-driving scenarios while leveraging a novel robust image processing pipeline to handle diverse environmental situations. Our approach relies solely on visual perception as the input to generate control commands. We trained and evaluated our methodology using a proprietary dataset from a self-driving car simulator consisting of image frames paired with their corresponding steering angles. The presented methodology is robust against overfitting, and it shows superior performance in terms of Mean Squared Error (MSE) and Mean Absolute Error (MAE) compared to previous methods.