Two or three course corrections do most of the work for a search drone

A search-and-rescue drone, mid-mission over a ridgeline, spots a heat signature and reroutes itself. A moment later it picks up a sensor blip and changes course again. That is enough. A new INFORMS Journal on Computing paper from UT Austin's McCombs School and the University of Michigan says two or three replanning rounds, meaning mid-flight reroutes triggered by new data, capture nearly all the benefit of a drone that replans every step. The payoff: 15 times the speed of a fully adaptive system, at a 12 percent increase in total route cost, and no measurable accuracy gain past the third round.

"You don't really need all the data to make good decisions," said Rohan Ghuge, a co-author. The work, also published at AISTATS 2024 and posted on arXiv, tests a hybrid plan. The drone fixes a route in advance, then adapts it only when it sees something worth chasing. The result is less onboard compute, longer mission range, and the same rough coverage as a drone that second-guesses every meter of flight.

The numbers come from computer simulations of search-and-rescue, infrastructure, and agriculture routes, not from drones in the field. The trade-off has not yet been tested in real flight.