Flying Drone Test Suggests Repetitive Crashing Is the Key to Self-Driving Succes

CMU robotics autonomous drone 4 photos
Photo: YouTube screenshot
CMU robotics autonomous droneCMU robotics autonomous droneCMU robotics autonomous drone
The way the automotive industry is going about the autonomous driving thing is by teaching a machine what it should do through preloaded algorithms and deep learning based on the inputs of a human driver.
Vehicles clad in video cameras, radars, and lidars are currently buzzing on the streets in all corners of the world trying to be the first to achieve zero-crash self-driving proficiency. But what if there was another solution?

A group of engineers thought about approaching the matter from another angle. They thought an AI might learn better how not to crash if it experienced crashing first. Lots of it. You know parents try to keep their offspring from making certain mistakes, but the little ones just don't listen and have to learn the hard way? This is sort of the same.

Evan Ackerman puts it even better on the Spectrum IEEE website: "One way to think of flying (or driving or walking or any other form of motion) is that success is simply a continual failure to crash. From this perspective, the most effective way of learning how to fly is by getting a lot of experience crashing so that you know exactly what to avoid, and once you can reliably avoid crashing, you by definition know how to fly."

11.500 crashes later, and the test drone used by the team at CMU robotics was able to fly by itself through indoor locations that kept changing their shape and avoid bumping into objects. It took 40 flying hours through 20 different environments for the drone to be able to not crash on its own, but it worked. What's more, it was also able to avoid featureless white walls and glass windows.

The drone has three cameras, one facing forward and two on each side. Whenever it encountered an obstacle and hit it, the image was saved so the next time it came across it, it would turn left or right, depending on which of the two seemed safer. If none of them looked particularly inviting, the drone would turn back.

Obviously, this is a drastic simplification of what cars have to deal with. Our vehicles need to navigate on crowded roads, deal with road markings, street signs and lights and who knows what else. Besides, crashing an inexpensive drone 11,500 times into a bookshelf is one thing, but doing it with a full-sized car is completely another. Not to mention the number of crashes necessary - due to the high number of variables out on the road - would be impossible to calculate. Still, what the guys at CMU robotics did is a nice exercise on how to look at a problem from a different angle.

If you liked the article, please follow us:  Google News icon Google News Youtube Instagram
About the author: Vlad Mitrache
Vlad Mitrache profile photo

"Boy meets car, boy loves car, boy gets journalism degree and starts job writing and editing at a car magazine" - 5/5. (Vlad Mitrache if he was a movie)
Full profile


Would you like AUTOEVOLUTION to send you notifications?

You will only receive our top stories