Key points are not available for this paper at this time.
Accurate and efficient multi-agent trajectory prediction remains a core challenge for autonomous driving, particularly in modeling complex interactions while maintaining physical plausibility and computational efficiency. Many existing methods- especially those based on large transformer architectures- tend to overlook physical constraints, leading to unrealistic predictions and high deployment costs. In this work, we propose a lightweight trajectory prediction framework that integrates physical information to enhance interaction modeling and runtime performance. Our method introduces two physically inspired strategies: (1) a constraint-guided mechanism is used to filter irrelevant or distracting neighbors, and (2) a physics-aware attention module is applied to steer attention weights toward physically plausible interactions. The overall architecture adopts a modular and vectorized design, effectively reducing model complexity and inference latency. Experiments on the Argoverse V1 dataset, comparing against multiple existing methods, demonstrate that our approach achieves a favorable balance among accuracy, physical feasibility, and efficiency, running in real time on a commodity desktop GPU. Future work will focus on validating its performance on embedded hardware.
Yu et al. (Fri,) studied this question.