INTERACTION-AWARE MULTI-AGENT BEHAVIOR PREDICTION

Abstract

Multi-agent systems are prevalent worldwide, from purely physical systems in basic scientific research to complicated autonomous systems in frontier technologies. Successfully predicting the agent's future behavior in these systems provides insightful foresight into potential outcomes and lays a solid foundation for advancing automated decision-making, thereby unlocking the potential for smarter, safer, and more efficient agents and systems. This thesis proposes a series of prediction methods targeting a fundamental research goal: achieving a precise, robust, and reliable behavior prediction with effective interaction modeling. Specifically, this thesis designs a series of methods focusing on three challenging aspects of multi-agent interaction: interaction comprehensiveness, interaction robustness, and interaction consistency. The proposed prediction methods adapt to various multi-agent systems and achieve effective prediction performance. Our main contributions include: 1) DynGroupNet: Proposes a dynamic-group-aware network, which captures and analyzes time-varying pair-wise and group-wise interactions with relational reasoning. 2) AuxFormer: Proposes an auxiliary learning framework to enhance the robustness of the prediction with interaction modeling data recovery capability and capture complex spatial-temporal interactions among body joints' coordinates. 3) EqMotion: Proposes an efficient equivariant behavior motion prediction model with invariant interaction reasoning to ensure the property theoretically.