KITNet: A Region-Attention-Activated Trajectory Predictor With Hierarchical Graph Neural Network in Dynamic-Mutant Multi-Agent System

Safe and efficient operation of Autonomous Delivery Vehicles (ADVs) in dynamic multi-agent environments, such as university campuses and industrial parks, necessitates accurate trajectory prediction of interacting agents. Conventional autonomous navigation systems, however, rely heavily on reactive real-time perception and often fail to predict complex spatio-temporal interactions among heterogeneous agents. This limitation frequently leads to suboptimal motion planning outcomes and operational inefficiencies, including deadlock situations, in congested scenarios. This paper introduces Knowledge-Interaction-Temporal Network (KITNet), a novel trajectory prediction framework specifically designed for ADVs operating in such complex, dynamic settings. KITNet employs a hierarchical Graph Neural Network (GNN) architecture to model intricate interaction dynamics, incorporating a novel attention mechanism based on set theory for enhanced spatio-temporal feature extraction and prediction of diverse behavior patterns. We evaluate KITNet on several trajectory prediction benchmarks according to the different tailored behavior modes under the defined mode space, including the ETH/UCY pedestrian dataset, the NGSIM highway driving dataset, and the Argoverse 2 urban driving dataset. Our results demonstrate state-of-the-art prediction accuracy, outperforming or matching existing graph-based and recurrent approaches. Furthermore, we discuss the integration of KITNet’s predictive outputs into local motion planning modules, showing potential for significantly reducing conflict scenarios and optimizing trajectory execution for ADVs. These findings establish KITNet as a highly effective trajectory predictor for autonomous transport systems, critically advancing predictive navigation and bridging the gap between perception and robust intelligent decision-making in complex urban and campus environments.