A research team has built an AI model that watches roadside traffic cameras and decides how to time traffic lights, though it hasn't touched a real intersection yet.
The model, called VLALight, is described as the first vision-language-action system built specifically for traffic signal control. Instead of relying on manually coded traffic states or a separate camera-analysis module bolted onto a rules engine, it reads multi-view roadside video directly and outputs signal decisions. The researchers trained it in two stages: a supervised "cold start" to teach basic visual and decision-making skills, followed by reinforcement learning that optimizes both individual intersections and the wider network. It also switches between fast and slow reasoning, spending more computation only when the payoff justifies the delay. Across seven real-world traffic-flow datasets spanning three cities, VLALight beat baseline methods built on traditional traffic engineering, reinforcement learning, and other language- and vision-based models.
The interesting part isn't the benchmark win, it's the framing: today's smart-traffic systems mostly still separate "seeing" from "deciding," which adds latency and brittleness at busy intersections. VLALight is a bet that reasoning straight from video, and coordinating that reasoning across neighboring intersections, produces smoother traffic flow than stitching together a camera system and a control algorithm after the fact.
That said, this is a paper with benchmark numbers, not a pilot program. Real intersections bring weather, sensor grime, and legacy hardware that historical datasets don't, and plenty of promising traffic AI has stalled right at that step.