UCLA

Autonomous racing requires precise raceline optimization to achieve faster lap times and efficient navigation. Traditional approaches, such as the K1999 algorithm, rely on geometric and physics-based modeling to determine optimal paths. This thesis explores the potential of deep learning to enhance raceline optimization by integrating the K1999 algorithm with modern neural architectures. Several models are evaluated, including Convolutional Neural Networks (CNN), U-Net, Transformers, and a hybrid U-Net+Transformer model. Among these, the combined U-Net+Transformer approach achieves the best performance in predicting optimal racelines. Evaluation extends to both AWS (Amazon Web Services) DeepRacer and F1Tenth (Formula One-Tenth) race tracks to demonstrate generalizability. The results highlight the effectiveness of deep learning in capturing complex spatial patterns, bridging traditional algorithmic techniques with data-driven methods for AI-powered autonomous racing.