FLAVOR mapping cuts drone exploration time by up to 42% and map memory by up to 99%

Researchers led by Yizhe Zhang have presented FLAVOR, a mapping framework that lets drones explore large areas with limited onboard memory and computing, in a preprint posted on arXiv.

Most drone exploration systems keep a global map that can be queried, so memory and query costs grow with the area explored, while maps centred on the robot forget the wider task, the authors write. FLAVOR keeps dense geometry only in a local map around the drone and stores compact records of what LiDAR keyframes saw, unexplored frontiers linked to odometry anchors and a sparse graph of routes between them. When the drone returns to a region, it rebuilds the frontier from those records and checks the route against the current local map.

In three simulated scenes, it explored in 31.98% to 41.88% less time and used 94% to 99% less map memory than the best baseline that finished all of them. In real-world use it cut the memory footprint by 92.67% in a forest and 62.42% on a construction site compared with an ikd-tree point-cloud map.