Goal-guided multi-agent motion prediction with interactive state refinement

Abstract

Motion prediction in the context of autonomous driving seeks to accurately forecast the potential future trajectories of various agents (i.e. vehicles, cyclists, and pedestrians) surrounding the autonomous vehicle. Enhancing the accuracy of motion prediction allows the autonomous vehicle to gain a more comprehensive understanding of its environment, thereby improving both driving efficiency and safety. Two key challenges in this research area are modeling the spatiotemporal interactions among agents and addressing the inherent uncertainty in agent intentions. To tackle these challenges, this paper presents a goal-guided multi-agent motion prediction framework with interactive state refinement (GISR). First, we propose a dual-branch spatiotemporal interaction modeling network that integrates graph neural networks and attention mechanisms to effectively capture the spatiotemporal relationships among agents. Second, to account for a diverse range of agent intentions, we design a query-based spatiotemporal fusion module, which employs a set of learnable goal queries to iteratively integrate knowledge-rich spatiotemporal features and generate reliable potential goals. Subsequently, we generate various plausible coarse trajectories associated with these goals, along with confidence levels for each modality. Finally, to ensure robust guidance for predicting the future states of interacting agents, we introduce an interaction-aware state refinement network that iteratively optimizes the coarse predictions by modeling future agent interactions, ultimately producing more realistic and socially acceptable trajectories. Experimental results demonstrate that GISR outperforms state-of-the-art methods on two large-scale motion forecasting datasets, Argoverse 1 and Argoverse 2, and exhibits strong capability in handling a diverse range of traffic scenarios.