At a junction in central Beijing, the red light no longer keeps a strict appointment with the clock. When the cross street is empty, the green opens early; when a queue builds on the main road, the phase stretches to let it clear. Behind this small change is software that is learning to read the road instead of obeying a fixed timetable.
The experiment is not unique to one city. Across China, transport teams are handing parts of signal control to models that watch traffic as it moves. The pitch is simple: a light should serve the cars in front of it, not a plan written months ago.
Key takeaways
- Beijing's traffic police reported pilot intersections saw vehicle speeds rise about 21% and the congestion index fall about 19% (People's Daily, Oct 2026).
- Shanghai has scaled AI signal control to 1,200+ intersections, 15 roads and 5 zones, with speed up 4.1–14.3% by layer (Xinhua, July 2026).
- Hangzhou's Xiaoshan district runs 1,260 junctions on one platform; main-road speed rose ~6.3% and congestion alerts fell 6% (China Economic Net).
- Hefei's Qianshan Road trial lifted average speed from 24.9 to 40.2 km/h and cut stops from 3 to 1 (China Mobile / People's Daily platform).
- The AI only tunes temporary plans; fixed timing and safety limits stay under human control, Beijing police note.
From fixed timetables to live reading
For decades, most urban signals ran on plans set by engineers: fixed cycles for morning peak, midday, and night, with small local tweaks. The method is predictable but blind to the moment. A rainy commute, a stadium emptying, or a blocked lane can undo a carefully drawn plan within minutes.
The newer approach uses a multimodal large model (大模型) as the coordinating brain. It pulls camera feeds, radar, and navigation data, forms a picture of the whole area, and proposes timing changes every few minutes. At night, when almost no cars are present, some junctions switch to on-demand control — a green appears only for the direction that actually has a vehicle.
What the pilots measured
The published results are modest but consistent across cities:
- Beijing: In a People's Daily account dated 17 October 2026, written by a Beijing traffic-police officer, pilot intersections recorded overall vehicle speeds up about 21% and a congestion index down about 19%. The same article said the city has roughly 12,000 signals, of which 7,900 are networked, with 100% networking inside the Fifth Ring and the sub-centre.
- Shanghai: Xinhua reported on 23 July 2026 that an AI anti-congestion agent had been deployed at more than 1,200 intersections, 15 key roads, and 5 core areas. Average speeds rose 4.1% at single junctions, 14.3% on road segments, and 12% in core zones. In one Yangpu district cluster around Changhai Hospital, covering 20 junctions over 1.7 km², speed rose 13.2% even as traffic volume grew 14.9%.
- Hangzhou: In Xiaoshan district, a cloud-based signal platform built by Amap (高德地图) and Insigma (中控信息) analyses 1,260 junctions and 3,600-plus road segments, generating 130 "green-wave" corridors. China Economic Net reported main-road speed up about 6.3% and congestion alarms down 6%.
- Hefei: On Qianshan Road, China Mobile's trial lifted average speed from 24.9 to 40.2 km/h and reduced stops from three to one, according to the People's Daily national media platform.
None of these claims a revolution. They describe single-digit to low-double-digit percentage gains on specific corridors, which is the realistic texture of applied traffic engineering.
How the system stays safe
Transport authorities are explicit that the model does not run the city unconstrained. Beijing's police officer noted the AI can only modify temporary plans and limited control modes; it cannot alter the underlying fixed timing. Every plan the AI proposes passes a compliance check, with hard limits on how far any green or red may shift.
This matters because a mistimed light is not a software glitch — it is a physical risk. Keeping a human-written baseline underneath the model is the difference between an assistant and an autopilot.
Where it spreads next
The build-out is still early. Beijing's pilot covers only parts of the city, though its networking foundation is broad. China Economic Net, citing IDC, projected that by 2028 more than 50% of major Chinese cities will use AI-enabled, software-driven traffic management to automate signal control. That is a forecast, not a deployment, and should be read as direction rather than fact.
A quieter trend may matter more: cities are wiring signals together first. Beijing reached 100% networking inside the Fifth Ring and the sub-centre, and ran dynamic "green-wave" algorithms on over 200 arterial roads in 2024, according to China Economic Net. Models are only as good as the network they sit on.
Honest limitations
The gains cited come from city traffic authorities and state media, often measured against their own pre-pilot baselines rather than independent audits. Percentage improvements vary by corridor and time of day, so city-wide averages are likely smaller than the headline figures. The underlying model quality, data sources, and failure modes have not been published in peer-reviewed form. Long-term effects on driver behaviour and on adjacent untreated roads are not yet documented.
What readers can do now
- City transport officials should publish raw before-and-after speeds per corridor, not just percentages, so gains can be independently checked.
- Residents can use open congestion indices (Baidu Maps, Amap) to compare claimed improvements against their own commute experience.
- Researchers should request anonymised signal-log data from pilot cities to test whether gains hold outside peak hours and across weather conditions.