New AI Helps Self-Driving Cars Stop at Red Lights

Self-driving cars learn by watching humans drive, but there's a problem: the AI almost never gets to practice stopping at red lights. Why? Because in thousands of driving examples, red-light stops are rare. So when the car finally faces one, it often creeps past the stop line. A new research paper from arXiv tackles this by reweighting training data — essentially making the AI practice those uncommon stops more often.

The method, called RedLight-VLA, adds two tricks. First, it emphasizes braking and accelerating moments in training, so the AI learns smooth stops instead of ignoring them. Second, it gives the model extra visual targets for traffic lights and stop lines — like a separate teacher pointing out where to look. Together, these help the AI connect what it sees with what it should do, without needing extra human rules.

In tests, the model reduced red-light stop-line overshoot from 7.3% to 6.8% and cut stop-line speed errors by 12.7%. Its predictions of where the car should be after three seconds improved by roughly 10%. There's a small trade-off: it now makes extra unnecessary stops at green lights 3.9% of the time, up from 3.2%. But the combined approach keeps that downside smaller than if you only used one trick.

For the rest of us, this brings autonomous vehicles one step closer to feeling human — smooth, predictable, and law-abiding. No more jerky roll-throughs. No more unsure hesitation at intersections. The paper is early research, not a production system, but it shows how we might get self-driving cars that actually respect traffic rules.