SWIFT: New AI framework boosts self-driving trajectory prediction with small-world networks

SWIFT, introduced by Chengyue Wang and colleagues, tackles a core challenge in autonomous driving: accurately predicting the future trajectories of vehicles, pedestrians, and other agents. Current data-driven methods often fail when traffic patterns shift (e.g., new city, different congestion levels) because they lack structural priors about how traffic actually flows. SWIFT explicitly encodes these priors using two innovations: a Small-World Interaction Network that captures both nearby interactions and long-range dependencies (like a car indirectly affecting another several lanes away), and a Flow Regime Encoder that adapts the interaction graph based on the current traffic state (free-flow vs. congested). A multi-relational graph module further distinguishes direct vs. higher-order agent relationships.

Experiments across three real-world datasets—nuScenes (city driving), MoCAD (mixed urban/highway), and NGSIM (highway)—show SWIFT consistently beats state-of-the-art baselines (e.g., Trajectron++, SceneTransformer) on standard metrics like Average Displacement Error and Final Displacement Error. Critically, SWIFT demonstrates strong generalization to unseen locations and traffic regimes, maintains accuracy when sensor noise is added, and achieves competitive performance with only 10% of the training data. The paper is accepted at IEEE TPAMI, signaling rigorous validation. For developers, this means safer and more reliable prediction modules that require less site-specific retraining—a key step toward scalable autonomous driving systems.