End-to-end system architectures in autonomous driving: Comparative analysis against modular design and technological exploration

Abstract. In the context of accelerated technological advancement, artificial intelligence is becoming an increasingly pivotal enabler of autonomous driving. With the integration of machine learning into autonomous driving systems, the traditional modular architecture has been disrupted, evolving into a more advanced end-to-end architecture. The end-to-end architecture has the advantages of simple structure and no cumulative error, especially under the current situation where a large amount of data and functions are continuously inputted, it demonstrates superior performance compared to the traditional modular system. However, the implementation of end-to-end is still mostly confined to academia and has not been widely applied to actual vehicles; the mainstream on actual vehicles is still the modular system architecture. As a result, the two system architectures are compared in this paper to ascertain their relative complexity, the extent to which errors are accumulated, their interpretability, the security they afford, and the extent of their data dependency. This is achieved through a comprehensive literature review. Furthermore, the paper delineates the deployment of the end-to-end system and outlines potential avenues for addressing the challenges that emerged during the implementation phase. The results indicate that the end-to-end system exhibits a certain degree of sophistication. However, a series of challenges pertaining to interpretability and security must be addressed to fully realize the potential of the end-to-end system.