Many of the applications for drones will see them spend most of their time high up in open airspace, but safely moving through denser urban areas at street level would be a handy capability, too. Researchers have come up with a control system for drones that enables them to autonomously navigate these busier settings, by showing them how cyclists and cars do their thing
This meant developing a deep learning algorithm they’ve dubbed DroNet. Instead of a suite of fancy sensors to observe the world around it, the algorithm relies only a regular camera similar to that of a smartphone to guide the drone safely around obstacles that might appear in its way.
Teaching DroNet to do this involved gathering mountains of training data, which the researchers captured from bikes and cars traveling through real-world urban environments. By showing DronNet over and over again how these vehicles navigate city streets, it eventually learned how to do so itself, figuring out how to move the drone down a road without crossing into the oncoming lane and how to stop when objects like pedestrians, construction works or other vehicle appear in its way.
“This is a computer algorithm that learns to solve complex tasks from a set of ‘training examples’ that show the drone how to do certain things and cope with some difficult situations, much like children learn from their parents or teachers,” says Davide Scaramuzza, Professor for Robotics and Perception at the University of Zurich.
The team also found that after being trained on urban settings, their drones could apply their learnings to other intimate environments like parking lots and office corridors. It says that there is further work to be done, but these kinds of developments edge us closer to a world where drones can zip through tree-lined streets to deliver parcels.
“Many technological issues must still be overcome before the most ambitious applications can become reality,” notes PhD student Antonio Loquercio, who worked on DroNet.
The video below shows how the system works, and team has published a paper online detailing its research.
Source: University of Zurich
Dr. Hans C. Mumm