DSIP uses diffusion models to replace traffic lights at intersections

A team of researchers from academia and industry (including authors from Shanghai Jiao Tong University and Ford Motor Company) introduced DSIP, a novel framework that leverages diffusion models—a class of generative AI—to coordinate connected and automated vehicles (CAVs) through intersections without any traffic signals. Traditional intersections rely on phase-based signals, forcing stop-and-go behavior that wastes time and fuel. DSIP shifts the paradigm to continuous, real-time trajectory optimization for all approaching vehicles simultaneously.

Using the SUMO microscopic traffic simulation platform, the team tested DSIP on various four-leg intersection layouts. The results show that diffusion-based coordination significantly outperforms both fixed-time signals and state-of-the-art reinforcement learning controllers in reducing average delay and increasing average speed, particularly under medium- to high-density traffic. The work isolates the theoretical upper-bound performance, assuming perfect communication and execution, to demonstrate the core benefits of the approach. This suggests that software-defined coordination can unlock latent intersection capacity without physical road expansion.