Hierarchical prediction uncertainty-aware motion planning for autonomous driving in lane-changing scenarios

Abstract

Autonomous vehicles (AVs) are expected to achieve safe and efficient interactions with surrounding dynamic objects. Multi-lane driving scenarios, however, intensify the complexity of AV navigation, given the uncertainties associated with neighboring vehicles’ lane-changing intentions and the subsequent travel trajectories. Deep learning has demonstrated effectiveness in unraveling complex motion patterns, enabling stochastic predictions of intentions and trajectories. Nonetheless, reduced performance of deep-learning-based prediction may be observed in unseen driving environments owing to their “black-box” nature, potentially leading to incorrect decision-making in AV navigation in these environments. To address these challenges, this paper proposes a comprehensive AV planning framework that integrates hierarchical behavior prediction via deep learning with motion planning based on dynamic programming. A set of safety criteria is introduced within the motion planning module to accommodate hierarchical uncertainties in behavior patterns, adjustable based on the reliability of the prediction model and eschewing rigid distributional assumptions. An improved constrained iterative linear quadratic regulator is devised to handle the corresponding non-convex constraints and to offer efficient online solutions for AV navigation. Evaluations conducted with the INTERACTION and HighD datasets demonstrate the effectiveness of uncertainty-aware planning in enhancing AV safety performance.