Multi-Agent Trajectory Prediction

Multi-agent trajectory prediction studies how to forecast several agents’ future movements together, where each agent’s actions may depend on others, the map, and the surrounding scene. Common approaches use graph neural networks, attention-based Transformers, recurrent or state-space models, and generative methods such as diffusion and flow matching to represent interactions and multiple plausible futures. Work spans driving benchmarks such as Argoverse and Waymo, pedestrian and robot navigation, and team sports, with methods increasingly incorporating uncertainty, physical constraints, and joint scene consistency. A central challenge is balancing accurate, diverse long-horizon forecasts with reliable social behavior, robustness to noisy or unfamiliar settings, and the low latency needed for safe planning.