Microsoft’s Experimental Self-Flying Aircraft Learn By Watching Hawks

The company is teaching gliders how to stay aloft without propulsion–and its research could inform a new generation of autonomous vehicles.

Microsoft’s Experimental Self-Flying Aircraft Learn By Watching Hawks
[Image: Microsoft]

Hawks rule the skies with a heavenly laziness. Simply by tilting their large wings, they can ride spiraling, hot convection currents up into the sky, gliding from current to current for hours on end. Each time a hawk leaves its warm thermal, it’s a cold dive into the unknown. It’s the world’s most soporific roller coaster. 


Such effortless flight requires a casual decision-making skill that Ashish Kapoor, principal researcher at Microsoft, admires–and no scientist truly understands. Birds of prey lack the comparatively massive gray matter of humans, yet they’re able to constantly weigh the risk and reward of leaving one hot air current for the next. It’s not so different from the decisions that must be made by modern-day AI systems. When should Google’s self-driving car choose to change lanes? How should an Amazon delivery drone approach a porch it hasn’t seen before? The answer may lie in animals that must make similar decisions without the brain power of humans. 

[Photo: Microsoft]
“This technology already exists, in forms of eagles, hawks, albatross. We know this is possible,” says Kapoor. “If [a bird] can do that with a peanut-sized brain, what would it require us to do, to build a technology that can reproduce those intelligent behaviors?”

To find out, Microsoft is building what one internal team has dubbed the “Infinite Flying Machine.” It’s really an entire fleet of sailplanes, also known as gliders, that the company has taught to ride thermals with zero propulsion. The goal is to build AI-driven robots capable of handling uncertainty, the same way a hawk handles the leap from one warm air stream to another.

It might sound silly, in an era when Uber, Tesla, and Google already have self-driving cars on streets, that Microsoft is teaching an aircraft the size of an RC hobby plane to fly. However, this robotic hawk makes a particularly good platform for Microsoft to test new AI–and to develop it without harming people along the way. “An autonomous car gone rogue is perhaps even more dangerous than a sailplane falling out of the sky where there’s no population,” says Kapoor. “With any robotic technology comes inherent risk, and it’s up to us as scientists and engineers is to manage the risk.”

For Microsoft, that means teaching a glider to ride thermals inside a simulation–and then turning it loose in the vacant skies of the Nevada desert.

[Photo: Microsoft]

Practice Makes Perfect, Even For Robots

Many of today’s AIs are trained by something called machine learning. By dumping massive amounts of data into a computer–like various pictures of a “tree”–and offering the right motivational feedback, the software will eventually learn to spot a tree even better than a human can.

The weakness with this approach is that it requires data. Lots of data. And the more abstract and dynamic the problem, the harder it is to build a big database of relevant information. Hot air thermals and other unpredictably real-world phenomena? Those are very abstract, dynamic problems with no easy training data set. “You need millions of examples for a machine learning system to get trained. And with real-world robotics, you don’t have that luxury,” says Kapoor. “When are you going to drive millions of billions of miles?”

Software simulators, with realistic physics just like a video game, offer one appealing alternative to real-world data when it comes to training AI. So before Microsoft put its glider in the real-life sky, it trained it to fly by watching hawks inside a simulator. The team built an open-source software called AirSim for its flight experiments, and over countless trials, various algorithms Microsoft developed learned how to fly like a hawk. There’s no onboard infrared camera to see thermal pockets and judge when to make the leap from one to the next. But a planning system on the ground predicts thermal conditions by accessing the local geography, while the drone’s own sensors include a GPS, barometer for a measuring altitude, pitot tube for tracking airspeed, and a system for gauging inertia.

Robotic simulators are a relatively new idea, but decades of video game development serve as a solid foundation for feasibility, and the approach is quickly gaining popularity. Google has used such simulators to train self-driving cars, and even to teach a machine how to walk with a humanoid body. Such simulators will no doubt be a battleground for the future of AI-driven robotics, as you can test theories without breaking hardware, and hone algorithms safely. That’s the overarching vision of Microsoft’s research here: that AI matures and learns within a simulation of the real world–until it’s ready to fly on its own. The company hopes to use this development pipeline at large in the future.

“For the sailplane in particular, we have bits and pieces of that pipeline working,” Kapoor says. Much of the development of algorithms was inside  a simulator. But it’s still evolving. It’s not just plug and play.”

[Photo: Microsoft]

To Build An Autonomous Car, First Build An Autonomous Plane

Now that it’s been trained in a safe simulator, Microsoft is testing the Infinite Flying Machine in real life. The planes feature various sensors, along with thermal cameras to spot hot currents. But otherwise, it’s an extremely simple piece of mechanical engineering that any radio controlled plane enthusiast would recognize. And that simplicity makes it a perfect platform to rapidly test AI.

“Cost is a big factor. Complexity is a big factor. I’m a pilot myself, as well as an amateur aircraft builder. I vividly remember the first time I saw a small plane. I was blown away by the simplicity of design. If you think about it, the aircraft is nothing but a canoe, controlled by a few cables to control the surfaces, and there’s a big fan at the nose,” says Kapoor. The Infinite Flying Machine is just as elegant. “All it’s got are three moving surfaces you control by pulleys. In terms of mechanical complexity, this is as simple as you can get. We’re pretty much flying a piece of carbon fiber or foam that you can twist and turn.”


With less risk of mechanical problems getting the way, Microsoft can focus purely on that software, making it better and better at weighing risk, balancing uncertainty, and thinking on the fly, so to speak. Already, its planes can catch drafts better than the team’s human pilots.

But a better drone is really secondary to Microsoft’s goal. The company believes the fundamental lessons of handling uncertainty learned by the drone could be applied to any AI robot. “The algorithms are pretty general in application,” says Kapoor. “They can be applied to many different vehicles.”

From the outside, it’s hard to know exactly how important these drones really are to Microsoft’s autonomous future. Kapoor sees them as a platform Microsoft can tap again and again to refine new AI software. And yet, Microsoft has already partnered with Baidu–often dubbed China’s Google–to share the search giant’s self-driving car technology developed abroad. It’s an illustration of how expansive many companies’ AI efforts are today. Rather than directing all of their resources at one big team with one direction, many companies are testing many different hypotheses all of the time.

But perhaps Kapoor’s confidence is born from a simple truth: that if there’s any certainty in the world, it’s uncertainty. And so there could be no more vital field of AI robotics research than mastering just that.

About the author

Mark Wilson is a senior writer at Fast Company. He started, a simple way to give back every day.