ADDI: A Simplified E2E Autonomous Driving Model with Distinct Experts and Implicit Interactions

Dapeng Zhang, Zhenlong Yuan, Chenyang Li, Yinda Chen, Shiyue Zhao, Hongtao Nie, Rui Zhou, Qingguo Zhou

Keywords: CV, Imitation Learning, Applications, 3D vision

TL;DR: A simple and efficient end-to-end autonomous driving method.

Abstract: End-to-end autonomous driving has emerged as a promising research trend aimed at achieving autonomy from a human-like driving perspective. Traditional solutions often divide the task into four sub-tasks—tracking-by-detection, online mapping, prediction, and planning—with several interactions to polish planning. However, this modular approach disrupts the cohesion of autonomous driving by ecomposing these processes and then linking them through interactions, leading to suboptimal and inefficient practical applications. To address this limitation, we propose ADDI, a simple and efficient end-to-end autonomous driving method. First, ADDI integrates tracking-by-detection and online mapping through a unified detection module paired with distinct expert designs, enabling simultaneous output of detection and mapping elements. Second, ADDI employs a unified motion planning model with distinct experts to jointly predict agent trajectories and ego planning trajectories. With this unified model structure, most interactions required by previous methods are rendered unnecessary. ADDI implements two implicit (resource-free) and two explicit interactions to associate the different components. Experimental results demonstrate that ADDI achieves state-of-the-art performance on both open-loop and closed-loop benchmarks while running significantly faster than prior end-to-end methods.

Primary Area: applications to robotics, autonomy, planning

Submission Number: 2357