A Comparative Analysis of End-to-End Motion Planning Architectures in Autonomous Driving

The software architecture of autonomous vehicles is currently fracturing into two distinct philosophies. For decades, the industry relied on a modular pipeline where perception, prediction, and planning occurred in strict sequence. Recently, deep learning has enabled endto-end architectures that collapse these distinct steps into unified networks. This study explores this transition by analyzing five representative methodologies that span the gap between traditional engineering and modern artificial intelligence approaches. The analysis examines the optimization-based integration proposed by Kedia et al. (2023), the unified deep learning models of Casas et al. (2021) and Hu et al. (2023), and the efficiency-focused architectures introduced by Weng et al. (2024) and Feng et al. (2024). By tracing the evolution from mathematical optimization to parallelized neural networks, this paper evaluates how each approach balances the competing demands of computational speed, interpretability, and safety. Our comparative analysis reveals fundamental trade-offs in autonomous driving system design and identifies promising directions for future research in hybrid architectures that combine the interpretability of optimization-based methods with the performance advantages of end-to-end learning.