Delivery Drones Can Learn to See, Dodge Obstacles In-Flight

Researchers have developed a method to enable aerial drones to learn to avoid in-flight collisions. The University of Zurich in Switzerland and Intel have jointly developed a method to enable aerial drones to learn to avoid in-flight collisions. Zurich's Elia Kaufmann and colleagues sought to develop drones that could self-pilot through hoops or gates, and set up a track with such gates arranged in a circuit, altering the obstacles' positions after each lap. Camera recordings of gate images gathered over several hours enabled a neural network to learn how to pass through them on appropriate trajectories, via generalization. Kaufmann says real-world testing showed the drone could pilot itself at up to about 3.5 meters a second.