Intermediate Tasks Enhanced End-to-End Autonomous Driving with Uncertainty Estimation

Autonomous driving in urban scenarios involves high-density dynamic objects and complex road conditions, requiring precise perception of both geometric and semantic information within the environment. In addition, the inevitable long-tail events also pose a challenge to safety. In this paper, we propose ITEUE, a novel end-to-end autonomous driving method which utilizes additional intermediate tasks to guide the learning process of the model. This help to capturing more traffic-related semantic and geometric information to enhance the representational capacity of the learned features and support proper decision-making. Additionally, an uncertainty-based method is employed to quantify the reliability of the model decision, contributing to the detection of latent long-tail adverse events and ensuring safety. We have conducted a series of experiments to compare ITEUE with previous works in complex urban environments on the CARLA simulator. The results demonstrate the effectiveness of ITEUE.