SG-SRC: Semantics-Guided State Residual Coupling for Dynamic Multi-Agent Trajectory Prediction

Abstract

In autonomous driving, heterogeneous agents interact asymmetrically and change dynamically. This often causes coordination inconsistency, policy mismatch and long-horizon prediction drift. The problem becomes more challenging when communication availability, interaction connectivity, and information freshness change dynamically over time. Frequent state sharing brings redundant information exchange, while delayed updates may weaken the modelling of evolving interactions. To address these issues, this paper proposes a Semantics-Guided State Residual Coupling (SG-SRC) framework for dynamic multi-agent trajectory prediction. In SG-SRC, a continuous-time state encoder is first used to model nonlinear motion evolution and capture fine-grained temporal variations in dynamic traffic scenes. Instead of transmitting complete agent states, compact liquid-state residuals are selectively exchanged. These residuals are designed to represent interaction-relevant semantic changes. A semantics-guided residual trigger is then constructed by jointly considering state deviation, rhythm mismatch and prediction uncertainty. Communication is activated only at critical interaction moments. In addition, a phase-coupled alignment mechanism is introduced to regulate asynchronous strategy rhythms. The received residual information is further fused by an attention-based liquid-state sharing module, which improves the consistency of multi-agent interaction modelling. Experimental results on the Argoverse1 and Argoverse2 datasets show that SG-SRC improves prediction accuracy, enhances interaction stability and reduces redundant communication in dynamic and non-stationary autonomous driving scenarios.