Two New Papers Tackle the Same Problem Traffic AI Keeps Ignoring: Reality

Two seconds doesn't sound like much until you're merging onto a highway at 60 mph. That's roughly 175 feet of travel on outdated information.

The second paper takes a different angle. Instead of fixing how AI handles bad data, it rethinks what the AI is actually trying to optimize. The Momentum-Based Reward Function approach argues that existing traffic signal RL systems are, sort of, asking the wrong question.

Typical reward functions penalize delay or queue length. Makes sense on paper. But the researchers found this produces "short-sighted or unstable policies." Their alternative: reward vehicles for keeping moving rather than just penalizing congestion. The distinction is subtle but apparently significant for learning stability.

I initially thought this was just semantic reframing, but after reading their results, I think there's something real here. They tested in SUMO (Simulation of Urban MObility, which seems to be the standard testbed for this stuff) and found better throughput-emission trade-offs than delay-based rewards or classical controllers like Max Pressure.

What Actually Matters Here

• Delay-awareness isn't optional. Any system that assumes instantaneous communication will fail in deployment. The DAROM paper shows you can build robustness in, but you have to actually try.

• Reward design shapes everything. The momentum paper is a good reminder that how you define success determines what your AI learns. Penalizing bad outcomes and encouraging good ones aren't always the same thing.

• Simulation results are promising but limited. Both papers use SUMO. Both show strong numbers. Neither has been tested on actual roads with actual humans doing unpredictable human things. This is worth remembering.

• The infrastructure dependency is real. DAROM assumes roadside units exist and work. That's true in some places, not in most. Scaling this requires hardware deployment, not just better algorithms.

You might be wondering why I'm covering two papers that seem pretty different. Honestly, I think they're two sides of the same coin. One asks: how do we handle imperfect information? The other asks: are we even measuring the right thing? Both questions matter if we want traffic AI that works outside of clean simulations.

The throughput-emission trade-off in the momentum paper is particularly interesting given where policy is heading. If you can show that your approach reduces CO2 alongside improving flow, that's a much easier sell to city planners than "we made cars wait 8% less."

I should note that I only found limited independent verification of these results. The DAROM paper has gone through five revisions on arXiv, which suggests active development (or, less charitably, that earlier versions had issues). The momentum paper is brand new. It's too early to say whether these approaches will hold up under broader scrutiny.

But tbh, I'd rather see researchers wrestling with real-world constraints than publishing another paper that assumes perfect conditions. That's progress, even if the solutions aren't fully baked yet.

Sarah Williams · 19 Jun · 3 min