EnTAIL: Evolutional temporal-aware interaction learning for motion forecasting

Accurately predicting the future trajectories of traffic agents in real-world scenarios is critical for advancing intelligent cyber–physical systems (CPS), such as autonomous driving systems and smart cities. A fundamental challenge lies in mining the evolving interaction patterns among multiple agents from their past trajectories, as traffic scenarios often exhibit complex interactions and continuously evolve along the timeline. However, existing methods fail to fully exploit the temporality inherent in sequential interactions. In the process of modeling interactions, they lack a comprehensive understanding of static interactions that occur at constant timestamps and the evolving patterns of interactions across timestamps. To tackle these challenges, we propose E volutio n al T emporal- A ware I nteraction L earning ( EnTAIL ), a novel temporal-aware interaction learning framework to model and reason the interactions among agents. EnTAIL captures both static interaction patterns at individual timestamps and temporal-aware interaction patterns across timestamps through a unified framework. Specifically, we introduce a trainable constant time encoding to integrate with the interaction modeling in each individual timestamp, which aims to capture the static interaction information. We propose a dynamic evolution encoder to model temporal-aware interaction features, enabling learning both short-term and long-term interactions within multiscaled observation windows. Besides, EnTAIL also considers the temporal feature in the prediction stage and models the long-range interactions ignored during the encoding phase. Extensive experiments conducted on the challenging real-world Argoverse dataset demonstrate that our proposed model achieves substantial performance improvement and outperforms the baseline methods up to 2.5% in minimum Average Displacement Error (minADE) and 1.2% in minimum Final Displacement Error (minFDE).