5G Cars Can Now Manage Their Own Radio Power — Using AI

Picture a busy city intersection at rush hour — not just clogged with cars, but humming with invisible radio signals. Every connected vehicle is broadcasting its speed, position, and braking status. Traffic lights push signal updates. Pedestrians' phones ping nearby cars for safety alerts. In a 5G smart city, all of this happens at once, constantly. Now Vellore Institute of Technology researchers have published a study in Scientific Reports showing that a lightweight AI agent — running inside each vehicle with no outside help — can manage this chaos more efficiently than any fixed system designed today.

Every car is already a radio station — and the airwaves are getting jammed

The technical name for what connected vehicles do is V2X communication — Vehicle-to-Everything. Car to car. Car to traffic light. Car to pedestrian. Car to the nearest 5G tower. At a busy intersection, dozens of devices are all transmitting at the same time, on overlapping radio frequencies, from positions that shift every second. That's interference. And interference, left unmanaged, forces devices to blast more power just to be heard over the noise — which creates more interference, which demands more power still. It's a feedback loop that wastes enormous energy and can degrade the safety-critical messages that V2X was built to carry.

Why fixed rules always lose against moving traffic

The standard approach to managing transmission power in wireless networks is to pre-program a set of rules. If signal quality drops below this threshold, increase power by this amount. It works, more or less, in environments that stay roughly predictable — a factory floor, an office building. But roads don't cooperate. A rule calibrated for a quiet 7am street fails badly at a congested motorway junction two hours later, where a dozen vehicles are all moving at different speeds, in and out of each other's signal range, bouncing signals off concrete overpasses and glass towers. Every prior approach to this problem — from centralized servers that try to coordinate all vehicles at once to deep learning models that require heavy computation — either can't react fast enough or demands too much hardware to be practical inside a real vehicle.

How does a car teach itself to use less power?

The team's answer is Q-learning — a stripped-down form of reinforcement learning that deliberately avoids neural networks and heavy computation. Each vehicle runs its own agent, which observes three things: how far away the receiving device is, how strong the current signal is, and how much interference is coming from nearby transmitters. It picks a power level, transmits, then checks the result. Good outcome — clean signal at low power — earns a reward. Poor outcome — signal drops below the minimum threshold — earns a penalty. The agent stores what it learns in a simple lookup table and, over time, stops guessing and starts making smart decisions. What's interesting about this design is what it doesn't need. No central server. No coordination with other vehicles. If a roadside unit goes offline or a base station loses coverage, individual agents just adapt to the changed conditions on their own, because lower incoming signal quality automatically triggers a search for better power strategies. The researchers ran 100 training episodes over a simulated urban intersection and watched the reward fluctuations steadily narrow — the classic sign that a learning agent is converging on a stable policy.

What the numbers actually showed

The results weren't a smooth, tidy upward curve — and that's actually what makes them credible. Energy efficiency peaked at transmission distances of roughly 35–40 meters and again at 65–70 meters, then dropped sharply beyond 70 meters. Past that range, keeping a clean signal simply costs too much power for the efficiency gains to hold. A similar shape showed up with signal quality: energy efficiency climbed as SINR rose from 5 dB to around 14–15 dB, hit its peak there, then fell off past 16 dB as the agent had to burn progressively more power to push the signal higher without proportional benefit. For power level itself, the sweet spot sat around 0.4 watts. Below that, signal quality was too poor. Above 0.5 watts, efficiency dropped as power consumption outran throughput gains. None of these numbers are magic constants — they reflect the specific urban simulation parameters the team used. But the shape of the relationships holds a more general lesson: in dense, mobile wireless environments, more power is rarely the answer. The agent learned that on its own.

What this still isn't — and what comes next

Honest caveat: this is a simulation. A well-constructed one, using real vehicular datasets for distance, traffic density, and signal strength, but still a model. Real 5G hardware in a real city is messier. Multipath reflections off irregular buildings, unexpected interference from non-V2X devices, weather effects on millimetre-wave signals — none of those are fully captured yet. The team also used tabular Q-learning, which keeps computational costs low but limits how finely the agent can represent complex states. As vehicle density and network complexity grow, that table might need to grow too, or give way to a more sophisticated approach. The researchers flag multi-agent reinforcement learning — where vehicles share what they've learned and cooperate rather than each working alone — as the obvious next frontier. Testing on actual 5G NR-V2X testbeds would follow. Neither of those is a small step, but the groundwork is there.

• Less power, not more: — The agent consistently found that moderate, precise power levels outperform broadcasting at full strength, which has direct implications for how 5G smart city networks should be designed.
• No central brain needed: — Each vehicle manages itself independently, making the system far more resilient to individual component failures than architectures that depend on a coordinating server.
• Distance has a hard ceiling: — Trying to maintain V2X communication beyond roughly 70 meters in dense urban environments costs disproportionate energy — a limit city planners should factor into roadside unit placement.

"This decentralized solution provides the advantage of scalability, network failure robustness, and smooth integration with 5G NR-V2X infrastructure — practical for real-time implementation." — Brindha S. et al., Scientific Reports, 2026.

Advertisement