New AI Lets Factories Track Machines Without Adding Sensors

A digital twin is a live computer copy of a real thing — a jet engine, a wind turbine, a water pipe — that updates itself with data so software can spot problems before they happen. The problem is that keeping the copy accurate usually means bolting sensors onto everything, and that gets expensive fast: hardware, batteries, wiring and the network bill to ship all that data back.

This new paper, from researcher Vincenzo Sammartino, proposes skipping most of those sensors. The trick is that ordinary radio signals already bouncing around a building or a city can be read for clues about their surroundings — like noticing a room feels different by how sound bounces off the walls. So one transmission does two jobs at once: it senses what's happening and sends that information to a nearby computer. A large AI model then runs on that nearby machine (rather than in a distant data centre) and refreshes the digital twin. The AI also learns on the fly, by trial and error, how to divide up power, bandwidth, memory and processing time — like a traffic controller who gets better at the job every shift.

In simulated tests on a 500-device industrial setup, the system updated its digital copies 38.7% faster and used 27.4% less energy than the best existing methods, while staying above 95% sensing accuracy and 30 frames per second of AI processing. Lower lag matters because a slow twin is a useless twin — if a pump is overheating, you need to know now, not in a minute.

The catch: this is a simulation, not a working installation, and it assumes fast "6G edge" networks that barely exist yet. The paper is also awaiting peer review. Still, the direction is clear — fewer gadgets, less wiring, smaller bills — which is good news for anyone who runs a factory, farm, power grid or city water system.