Drones With Tiny AI Brains Beat Central Control in Rescue Tests

Imagine ten rescue drones sweeping a collapsed building. Today most drone swarms depend on one central computer telling each drone what to do. This team instead gave every drone its own small AI — a slimmed-down version of the technology behind chatbots — so each can think and act alone. The catch: when ten AI brains all shout at once, the radio links jam and each drone drowns in messages it does not need.

The researchers solved this with a sort of polite group chat. Each drone stores what it learns as short tagged notes — things like "survivor spotted here" — sorted into three piles: facts everyone needs, facts only it needs, and facts to tell specific neighbors. A simple rule-based messenger then shares a note only if it is new and genuinely useful to that particular drone. No AI decides who hears what; rules do.

In a simulated search-and-rescue mission with ten drones, this approach finished all runs. Drones that broadcast every message finished only 70 to 85 percent of runs. Letting the AI itself choose what to forward failed every single time. Their method also used roughly half the AI computing power, sent less data, and came closer to the true number of survivors.

The lesson reaches beyond drones: any fleet of AI helpers — delivery robots, warehouse machines, self-driving fleets — works better when it shares selectively rather than loudly. One honest caveat: this is a computer simulation with ten drones, not a real disaster site, so bad weather, weak signals, and much larger swarms still need testing.