Decomposing Trajectory Forecasting into Route and Motion: Enhancing End-to-end Autonomous Driving with Route Map Inputs

Abstract

Autonomous driving systems operating in urban environments require precise perception, planning, and accurate control to navigate complex traffic scenarios while respecting traffic rules and ensuring safety. Recent advancements in learning-based end-to-end approaches have showcased remarkable performance in goal-directed navigational scenarios. However, while most state-of-the-art approaches primarily focus on enhancing the perception module for a better understanding of the environment, they adopt a simple GRU-based autoregressive structure for producing waypoints. In this paper, we introduce RM2: Route and Motion prediction with Route Map, a simple yet effective approach that decomposes waypoint prediction into two distinct concepts: route, representing the future path to follow, and motion, determining the trajectory and speed. Through experiments, we discover that this approach is the most effective, leading to superior route completion in closed-loop evaluation. We also demonstrate the benefits of incorporating past route predictions. This way, the RM2 approach significantly outperforms the second-best choice by 50% in the lane-change benchmark routes newly introduced in this work.