Skip to main content

A tiny startup has made big strides in creating self-navigating drones

A tiny startup has made big strides in creating self-navigating drones


Veterans of Google X and MIT have created some remarkable computer vision systems

Share this story

Some of the biggest news out of CES this year was the huge progress companies have made on technology that can help drones be aware of their surroundings, allowing them to sense objects and avoid collisions. Those breakthroughs were powered by well-established drone companies like Ascending Technologies and semiconductor giants like Intel. Earlier this week, in a small park along the San Francisco Bay, I got to see another take on "sense and avoid." It came from a small startup, Skydio, and it was as impressive as anything I saw from the titans of tech.

Skydio was created by three friends who met at MIT. Adam Bry and Abe Bachrach were in the robust robotics program, researching ways to build aircraft that could fly themselves without GPS, culminating in a fixed wing drone with a laser range finder that autonomously navigated its way around a parking lot. After that, Adam and Abe were part of the founding team for Google X’s Project Wing, designing and building the delivery drone that flew over Australia.

After a year and a half, the pair left Google X and along with Matt Donahoe, who they met at MIT’s Media Lab, set out to create a system to power self-navigating drones using only the commodity chips and and sensors you would find in your average smartphone. Today Skydio, which has been working in stealth for around a year, announced a $3 million round of funding from Andreessen Horowitz and Accel.

Right now most drones stabilize themselves using GPS and some can even use that signal to navigate a course or follow you around. But they are helpless if they lose GPS and also can’t see trees on a ski slope and know to avoid them when they following you to film. Powerful sense and avoid would make drones safer by giving them the ability to correct for bad pilots. It would also make autonomous flight more robust, allowing drones to avoid obstacles. Without it, the fantasy of a delivery drone landing on your lawn would remain just that.

"The goal is to take something that normally costs $5,000 and sell it for $50."

At CES we saw the beginnings of this, using powerful, expensive sensors custom built by Intel. Skydio is betting that they can achieve robust sense and avoid using only ordinary cameras, instead of the lasers, sonar, and depth of field cameras we saw at CES. "The goal is to take something that normally costs $5,000 and sell it for $50," says Bry, the CEO. "We think vision is going to win the day. It’s an incredibly rich data source, it’s just algorithmically challenging. But computers are getting faster and these algos are coming into place."

Skydio gave us a demo of their drone, a prototype quadcopter with two small cameras feeding a video stream into a small Intel media center. The unit had no GPS, which is how a normal drone would orient itself, but after takeoff it hovered completely in place, holding its position as well as or better than any drone I’ve seen. "It’s doing that 100% based on the visual it has of us standing here," Donahoe explained.

The first demo Donahoe walked me through was called the "Magic Wand." You point your phone at the drone, which is connected to it over Wi-Fi. Tap to initiate control, then move your arm in any direction and the drone will move as well, pivoting to keep you in focus. It worked remarkably well and had a much simpler interface than any of the "follow me" drones I tested at CES.

Right now a bulky controller is the command prompt of drones

"Right now this is the command prompt on a drone," said Bry, hefting the RC flight controller, with its two joysticks and multitude of buttons and switches. "If you’re an expert operator, and you’re trained, you can get really incredible photos and video. We want to find a way to deliver that same power, but in a much more intuitive way that is open to anyone with a smartphone."

Since it’s based on computer vision, Skydio’s system would struggle to work under poor lighting conditions or in weather like rain and snow. "It’s kind of intuitive," said Bachrach. "If a person watching the video would have difficulty figuring out what’s going on, then the software probably would as well."

"It will seem crazy that a drone would ever crash."

I asked the Skydio team how good the system would have to be in order to be commercially viable. A drone that only crashes one in 10,000 times? One in a million? "Realistically the first systems won’t be perfect," says Bry. "But right now there is nothing available to the average consumer. Something is better than nothing."

The idea is to use sense and avoid technology initially as a redundant safety feature that helps pilots avoid obstacles. As it gets better— and assuming the FAA allows the consumer drone trend to continue — it will be employed in more autonomous situations, like filming your kid on the soccer field or checking out a leaky roof to see what needs a fix. "It’s going to be a very long term project to make a perfect, robust system, but I think in five or six years you will just trust the unit," says Bry. "It can see the world and know where it needs to move or land to avoid problems. It will seem crazy that a drone would ever crash."

Accel is also an investor in Vox Media, parent company of The Verge.