Researchers say a diffusion model can do a better job managing intersections than a traffic light — and the simulation results are hard to dismiss.
A paper published on arXiv introduces DSIP, short for Diffusion-model-based Signal-free Intersection Planner. Instead of cycling vehicles through fixed green-yellow-red phases, DSIP treats intersection management as a continuous trajectory optimization problem across multiple vehicles simultaneously. The system was tested using SUMO, a standard traffic simulation platform, across various four-leg intersection layouts. In medium- to high-density traffic, DSIP outperformed both conventional fixed-time signal control and reinforcement-learning-based controllers on average delay and average speed.
The appeal here is architectural. Traffic signals are a blunt instrument — they divide time into phases and hope the math averages out. DSIP instead plans paths for all vehicles together, filling gaps that phase-based systems structurally leave empty. If it holds up outside simulation, cities could squeeze more throughput from existing intersections without rebuilding them.
The catch the researchers are upfront about: results assume idealized communication and execution conditions. Every connected vehicle talks perfectly to every other, latency does not exist, and no one cuts anyone off. Real urban traffic is none of those things. Diffusion-based planning is also computationally heavier than a timer on a pole — deploying it at scale would require infrastructure that most cities do not have and connected vehicle fleets that remain years away from critical mass.