New Trick Teaches Drone Swarms to Fly Together on the Cheap

Getting drones to fly as a coordinated team — think search-and-rescue sweeps or crop-spraying fleets — usually requires training their software inside a simulated world that mirrors real physics almost perfectly. That's accurate, but brutally expensive. Each drone you add multiplies the computing work, and simulated drones crash into each other far more often, wasting time before anything useful is learned.

The new approach, from researchers Maxim Mednikov and Oren Gal, flips that. Instead of training in the expensive simulation, it trains the team in a rough, stripped-down one that treats each drone as a simple point in space — fast and cheap to run. The rough simulator is then corrected by a small patch learned once, in advance, from short calibration flights of a single isolated drone. Because it only takes one drone to calibrate, the data-collection bill doesn't grow as the team gets bigger.

On four cooperative flying tasks with teams ranging from 3 to 18 drones, the corrected policy beat the uncorrected rough version every single time, and beat a fully realistic simulator-trained policy in 22 of 24 matchups. It came close to a more heavily refined realistic training run, with the gap shrinking as teams got larger. Crucially, it skipped the sky-high crash rates that plague realistic training.

So what does this mean for you? Drones doing real work — inspecting power lines, spraying fields, mapping disaster zones, delivering packages — depend on software that's costly to develop. Cutting that cost could put capable autonomous fleets in reach of smaller companies and public agencies. The honest caveat: this was all tested inside computers, not in real wind, rain, or crowded airspace, so real-world proof is still ahead. The correction is also tuned to one drone design, so swapping hardware means recalibrating.