Scientists Built a 200,000-Mission Test Course for Search Drones

A team of researchers has released AerialDojo-200K — a giant practice ground for drones and other AI-powered flying robots. Think of it as driver's ed, except it exists entirely in simulation and covers 42 different 3D worlds: city streets, forests, bridges and power plants, plus disaster zones like collapsed buildings. In each world, a drone must find a specific object after being told what to look for — either in words ("the red truck") or with a reference photo.

That sounds simple. It isn't. Until now, drone research ran on small, homemade test sets that couldn't be compared fairly, which slowed progress. This new suite is roughly 3 times bigger in scenery and 18.7 times bigger in tasks than the previous best. It contains 205,732 individual missions. To build it, 12 human annotators spent two months labelling 109 landmarks and 2,099 target objects by hand. Each mission also includes a safe, collision-free flight path and video footage, so AI models have an example to learn from.

Then came the reality check. The team tested nine leading AI models — five open-source, four commercial — and none came close to being a general-purpose flying assistant. The hardest test involved 21 "out-of-distribution" scenes: places that look nothing like the practice data. Models that memorized familiar city layouts fell apart in unfamiliar terrain. The paper's own conclusion: there's still a long way to go.

Why you should care. A drone that hunts for a lost hiker, spots a gas leak, or grabs a specific box on a warehouse shelf would be genuinely useful — and a real business. This work doesn't deliver that tomorrow. It mostly measures how far off it is, and gives everyone a shared yardstick. Expect more capable drone autopilots in a few years, not a few months.