Skip to main content

Watch these A.I. drones teach themselves how to fly through trial and error

You know the saying, “If at first you don’t succeed, try, try again?” Well, it also counts for drones. At least, that is the takeaway message from a recent paper titled “Learning to Fly by Crashing,” published by roboticists from Carnegie Mellon University.  They subjected hapless drones to 11,500 collisions in 20 different indoor environments, spread over 40 hours of flying time, to prove it.

They did it for a good reason, too — and it is not because they have a whole lot of old quadcopters to get rid of before the start of the next academic year.

“We are interested in the problem of drone navigation: How does a UAV learn to avoid obstacles and learn to navigate,” Abhinav Gupta, an assistant professor in CMU’s Robotics Institute, told Digital Trends. “Unlike most other problems where data is the answer to many hard questions, what makes this problem hard is [a] scarcity of relevant data. We can use human experts and ask them to fly drones, but such data is small in size and biased towards success since the number of crashes is very low.”

Instead of using a computer simulation to solve the problem, Gupta and colleagues set out to build a framework where the goal of the drone is to crash. In their study, the drones were instructed to fly slowly until colliding with something, after which they would return to the starting position and set off in a new direction. By doing this repeatedly and then feeding the crash data into a convolutional neural network, the team was able to train a drone to be able to more successfully fly autonomously — even in narrow, cluttered environments.

The algorithm controlling the drone works by splitting the picture the drone sees into two separate images and then turning in the direction of whichever looks less likely to result in a crash. The results were surprisingly effective.


The drone still runs into problems, particularly involving glass doors and plain walls, but it is a whole lot better than it was before its training. Should we wind up living in a world where thousands of drones are constantly buzzing around, carrying out a range of tasks in complex real-world environment, research like this is going to be vital to developing better autonomous flying machines.

In the meantime, researchers get to exercise their destructive whims by making robots crash for the sake of “science.”

Editors' Recommendations

Luke Dormehl
I'm a UK-based tech writer covering Cool Tech at Digital Trends. I've also written for Fast Company, Wired, the Guardian…
IBM’s Wimbledon-watching A.I. is poised to revolutionize sports broadcasts
Tennis players competing at Wimbledon

Among the most lauded essays ever written about the game of tennis is David Foster Wallace’s 2006 article “Roger Federer as Religious Experience.” Originally appearing in the New York Times, the approximately 6,000-word tribute to one of the world’s most supremely talented players reads, as its title makes clear, more like a divine celebration than a piece of sportswriting.

Wallace (and he was certainly not the first writer to do this) gushed about high-level sporting achievements as though they were more than just superb technique; as if they were, somehow, a transcendent portal to godliness. Ordinary mortals like you and I could comprehend what was happening, but only barely. In order to truly appreciate Federer's athletic feats, we needed a member of the priesthood -- a talented youth player like Wallace had been -- who could make it intelligible to us.

Read more
A.I. can spot galaxy clusters millions of light-years away
ai identify galaxy clusters 205077 web 1

Image showing the galaxy cluster Abell1689. The novel deep learning tool Deep-CEE has been developed to speed up the process of finding galaxy clusters such as this one, and takes inspiration in its approach from the pioneer of galaxy cluster finding, George Abell, who manually searched thousands of photographic plates in the 1950s. NASA/ESA

Galaxy clusters are enormous structures of hundreds or even thousands of galaxies which move together, and for many years they were some of the largest known structures in the universe (until superclusters were discovered). But despite their massive size, they can be hard to identify because they are so very far away from us.

Read more
A.I. is getting scary good at generating fake humans. Watch this demo
japanese gan people dont exist screen shot 2019 05 03 at 16 23 06

【データグリッド】全身モデル自動生成AI | [DataGrid] Model generation AI

A.I. is getting pretty scarily good at lying to us. No, we’re not talking about wilfully misleading people for nefarious means, but rather creating sounds and images that appear real, but don't exist in the real world.

Read more