Researchers have used the power of artificial intelligence to develop a new navigation system for autonomous drones. Through deep learning, devices can avoid obstacles in complex environments at up to 40 km / h.
Drones must navigate with very limited computing power, andimperfect. Noise, motion blur, and differences in lighting make calculating a route very complicated. The researchers therefore called on (deep learning), and more precisely a convolutional neural network. They trained their algorithm in with various routes, and each time an optimal path to which the virtual devices had to approach as closely as possible.
Presentation of the autonomous flight system (activate automatic translation of subtitles). © UZH Robotics and Perception Group
An algorithm that works without being modified for real conditions
The researchers then launched the drones into the real world, with the algorithm obtained in the simulations left as is, without any adaptation. The devices were able to follow the course avoiding all obstacles at aof 18 km / h. From 25 km / h they see the first , but even at 36 km / h the drones manage to complete the course without hitting any obstacle in 60% of cases. Most collisions are due to narrow passages, less than a meter wide.
The advantage of this system is that it is even able to avoid moving obstacles, as long as they move slowly relative to the drone. Researchers say this new navigation system can’t yet beat human pilots, but it comes close.