A peer-reviewed ACL 2026 paper introduces Traffic-R1, a compact 3-billion-parameter model for running AI traffic lights. The authors report that the system is already in production on signals serving more than 55,000 drivers a day, that average queues are more than 5% shorter, and that operator workload is about half what it was.
Those production figures come from the authors themselves. They are a serious claim in a peer-reviewed conference paper, not an independent city’s audited before-and-after study, and they should be read that way.
In production, the model now manages signals affecting over 55,000 drivers daily, reduces average queue lengths by more than 5%, and halves operator workload.
If those numbers hold up under outside scrutiny, the everyday payoff is simple: a shorter wait at the light, and less frantic work for the people who keep the signals running.
What the system is
The paper is by Xingchen Zou, Yuhao Yang, Zheng Chen, Xixuan Hao, Yiqi Chen, Chao Huang, and Yuxuan Liang. Traffic-R1 was trained with expert guidance in a simulated traffic environment, then tested against strong baselines, including training-heavy reinforcement-learning controllers.
The authors highlight three design aims. The model is compact enough, they say, to run in real time on modest hardware at the roadside. It is meant to generalize to new road networks and unusual incidents without being retrained from scratch. And it is designed to explain its choices and coordinate across intersections.
For drivers, the technical details matter less than whether the light stays greener when a queue is building, and whether the system can cope when a crash, a concert, or a stalled bus breaks the usual pattern. Traditional timers follow a fixed plan. A model that can reason about the current queue, and talk to neighboring lights, might do better when the day refuses to follow the plan.
Why this matters for people
Sitting in traffic is time you do not get back. A system that shortens average queues by even a few percent, across tens of thousands of daily drivers, would add up. Operators, meanwhile, could spend less time firefighting routine timing and more time on the odd cases that still need a human.
Limits to keep in view
The production results are the authors’ own figures. The abstract does not provide an independent evaluation, and the source date is around July 2026. Cities differ, and a model that helps in one network may need careful checking in another. “Halved workload†and “more than 5% shorter queues†are encouraging if they are measured consistently, but outsiders will want to see the comparisons, the time window, and whether other changes on the road happened at the same time.
Benchmarks in simulation are not the same as a year of messy weather, roadworks, and human driving. The authors plan to release code and a model checkpoint, which should make follow-up work easier. Until cities or independent engineers publish their own measurements, the honest summary is that a peer-reviewed team reports shorter queues and a lighter workload, at a meaningful daily scale. That is worth close attention from cities, and it is not yet the last word on what the system does in the wild.
Looking ahead for AI traffic lights
The people-first hope is a shorter, calmer trip through an intersection, without asking a city to install an enormous new computer in every cabinet. Traffic-R1 is presented as a step in that direction, already claimed to be running for tens of thousands of drivers.
Open code will help other researchers check the method. Open city data on actual queues, travel times, and operator hours would help even more, because that is how ordinary drivers actually experience those same familiar traffic lights each and every ordinary very busy working day.
You can read the paper in the ACL Anthology. Drivers do not need a new jargon term for that result. They will notice if the line moves, morning after morning.
If independent city data later confirm the shorter queues and the lighter workload, more drivers may feel the benefit at the same red lights they already know too well.