teaching a team of drones to map a flood together
After a flood, rescue teams all ask the same first question: which buildings are underwater, and which are safe? Walking the streets to find out is slow and dangerous. A small team of drones can answer it from the air. This project teaches them to do it as a team, with no pilot and no one dividing up the map for them.
These are real neighbourhoods and real floods, from satellite imagery in the SpaceNet 8 dataset: a river valley in Germany during the July 2021 flood, and four areas of Louisiana after Hurricane Ida. Pick a map, pick a team, and launch. From there the trained AI is in charge.
Every mission here was flown by the trained network in the research simulator, under full physics: wind, draining batteries, limited charging bays, radio range and cameras that sometimes need a second pass. The flood itself is alive too: the water keeps rising while they fly, so a building checked safe early can go under later, and the team has to notice and go back. The team even decides for itself when the job is done: it never knows the true number of flooded buildings, and it may call the mission over early if it believes it has found them all. In every mission recorded here it chose to keep flying until it had checked every building. The flights were recorded move for move; you are replaying the team's actual decisions, including their mistakes. The faint lines show where each drone has decided to go next.
Small practice, big missions. The team trains on small practice maps, then searches areas hundreds of times larger than anything it saw in training, with up to 32 drones at once.
Tested on real floods. The final exam is not an invented town: the test maps are built from satellite images of real flooded neighbourhoods, the same ones you can fly above.
The simple way breaks in the real world. In an easy, idealised world, a simple pre-planned route is a genuinely strong rival, and early versions of this project said so. Then the simulator got the real world's complications: wind, batteries that drain, water that keeps rising mid-mission, cameras that sometimes get it wrong. Under those conditions the pre-planned route finished only two of the five real missions, and the classic back-and-forth sweep finished none. The learned team finished all five, brought every drone home, and missed less than one building per mission on average.
The project began as my master's research on drone coverage, published at the ICUAS 2024 conference, and has kept growing since.