Drones have a habit of . If they are ever to be relied on for delivering packages in complex environments like cities, they’re going to have to get smarter. A team of researchers from the University of Zurich and Intel has come up with –Ěýlearn to dodge obstacles as they fly.
and colleagues wanted to develop drones that could autonomously pilot themselves through hoops or gates used in drone racing. They prepared a track with such gates laid out in a circuit, but planned to quickly alter the obstacles’ positions after each lap. Would their drone be able to stay on course?
First, they manually carried the drones through an example track so that the device’s on-board camera could gather images of the gates as it passed through them. After two hours of this, the team had collected tens of thousands of images. The pictures allowed a neural network to learn how to pass through gates on appropriate trajectories – when to turn left or right, for example.
Advertisement
[video_player id=”S3h3A8zX” access_level=”everyone”]
“We didn’t have to program our drones, we just showed the drones what we wanted them to do,” explains Elia Kaufmann. “The network is able to generalise – if it gets new images that it has never seen before it is able to fly accordingly.”
In simulated experiments the system performed well so the team . The drone was able to pilot itself at speeds of up to about 3.5 metres per second – but the teamĚýreckons it could go much quicker in an environment where it had more space.
“It’s a key development,” says , director of the Aerial Robotics Laboratory at University College London. He adds that in certain contexts, having a speedy drone that can autonomously dodge things in their way could be very useful for making deliveries.
“For urgent delivery – such as in environments that have a lot of obstacles like forests, disaster zones, or inside collapsed buildings,” he says.
Journal reference:Ěý
