Yaw-guided end-to-end imitation learning for autonomous driving in urban environments

Existing imitation learning methods, such as CIL, face significant limitations in data utilization and generalization ability when addressing the road-option problem in urban environments. These methods often struggle with insufficient parameter optimization in branch networks and fail to adapt to dynamic and complex scenarios, such as intersections with dense traffic. To overcome these challenges, we propose Yaw-guided Imitation Learning with ResNet34 Attention (YILRatt), a novel end-to-end autonomous driving framework that leverages yaw angle guidance and an attention mechanism to enhance sample efficiency and adaptability. YILRatt utilizes yaw information derived from navigation map trajectories, eliminating the need for HD maps and enabling fully end-to-end operation with consumer-level GPS receivers. The integration of ResNet34 and the attention mechanism ensures accurate perception and provides interpretability through attention heatmaps, which reveal causal relationships between decision-making and scene perception. Experimental results on the Carla 0.9.11 simulator, including improved benchmarks CoRL2017 and NoCrash, demonstrate that YILRatt achieves a 26.27% higher success rate than CILRS. This improvement is particularly evident in dense traffic scenarios, where the attention mechanism effectively captures dynamic obstacles and enhances navigation performance. By addressing the limitations of existing methods, YILRatt offers a robust and interpretable solution for autonomous driving in complex urban environments.