Aston University researchers tap AI to kill traffic jams

Anyone who drives has experienced the rage of sitting at a bright red light while cross-traffic is completely empty.

In 2021 alone, Americans wasted 3.4 billion hours stuck in traffic, according to connected car analytics firm INRIX. That comes out to 36 hours per person, basically tossing an entire 40-hour workweek straight into the tailpipe.

Now, AI researchers and startups are stepping up to retire rigid, outdated traffic lights once and for all.

Moving past rigid timing charts

Most traffic systems today still rely on human engineers manually programming phase changes, or at best using magnetic road loops, cameras, and LiDAR to count cars. While firms like Tel Aviv's Intelligent Traffic Control and Texas-based Paradigm Traffic Systems are already using camera feeds to tweak signals, researchers at the UK's Aston University are pushing the envelope.

Researchers Maria Chli and George Vogiatzis built an AI traffic control system that Aston University claims is the first of its kind. Instead of relying on rigid, pre-programmed schedules, their system trains inside traffic simulations using deep reinforcement learning.

By playing through millions of virtual trial-and-error scenarios, the system learns how traffic actually behaves. Once deployed at a real-world intersection, it can regulate signals in real time and handle congestion patterns it has never encountered before.

A smarter brain on existing asphalt

The tech combines advanced computer vision with predictive machine learning. Sensors track the speed and trajectory of cars alongside pedestrians and cyclists, while CCTV feeds give the AI a bird's-eye view of the broader road network.

According to Chli, deep reinforcement learning processes this firehose of sensory data while factoring in complex variables that humans overlook, such as weather conditions and local air quality metrics.

Crucially, Vogiatzis noted that the platform works entirely on top of existing cars, roads, and traffic hardware. Cities do not need to tear up pavement or wait for every driver to buy an autonomous vehicle.

Why it matters

For everyday drivers, this could mean an immediate end to senseless red-light idling and wasted fuel, with zero need to buy a new car or install an app.

For city planners and municipal tech vendors, Aston University's work highlights a critical pivot from rigid rule-based automation to adaptive, generative reinforcement learning. Rather than spending billions on dedicated vehicle-to-infrastructure (V2X) hardware, cities can extract major efficiency gains by putting smarter software behind legacy cameras and lights.

We think training the model in high-fidelity simulations before hitting real asphalt is the smartest move here, eliminating dangerous real-world learning curves. However, the catch will be edge-case reliability. If a sensor gets blocked by snow or a bizarre multi-vehicle wreck scrambles the visual feed, the system must fail gracefully without paralyzing an entire downtown grid.