The Infinite Soaring Machine: AI Piloted Glider
Over the desert sands of Hawthorne, Nevada, the Microsoft glider dubbed “the infinite soaring machine” sails gracefully through the air without aid of engine, propeller, or even a human pilot.
Instead, this advanced AI-piloted craft stays aloft by actively predicting and seeking out pockets of rising air known as thermals, just as a birds of prey do.
“Birds do this seamlessly, and all they’re doing is harnessing nature,” said Ashish Kapoor, one of the principal researchers of the project. “And they do it with a peanut-sized brain.”
The thermals on which sailplanes ride are formed in three principal ways. First, in what is known as hill lift, wind blowing past a hill is forced up by the contour of the land. By flying perpendicular to the slope, a glider, can ride the wind to a higher altitude without expending any energy.
Second, in thermal lift, uneven heating by the sun raises the temperature of one section of ground more than the surrounding sections. This section of ground in turn heats up the air above it and creating a hot column of rising air that continues to pull more air up like a siphon.
Finally, wave lift is characterized by air moving in an undulating pattern after passing over a hill and rebounding off the ground with sufficient momentum. These are usually accompanied by wave clouds, which are eyebrow-shaped clouds that form at the peak of each wave.
(Diagrams courtesy of The Scottish Gliding Centre.)
By searching for telltale markers such as hills facing into the wind and wave clouds, a seasoned glider pilot, a bird of prey, and Microsoft’s AI can find thermals and ride them effortlessly. However, while the pilot uses logic and the bird uses instinct, the AI utilizes an altogether different method to spot thermals.
To start, the Microsoft sailplane is outfitted with a myriad of sensors to measure factors such as location, altitude, airspeed, temperature, pressure, wind direction and speed, and the lay of the surrounding terrain. The information gathered by these sensors is first sent to the high-level planner, which is responsible for predicting where thermals are most likely to occur and determining which direction to head in. It does this using complex algorithms such as the Markov decision process, which is designed to make decisions with only limited information, and Monte Carlo tree search, which helps to determine the most rewarding course of action – all in real time.
Once the high-level planner has “decided” where to go, it passes that information onto the low-level planner. Utilizing relatively simple Bayesian reinforcement learning, the low-level planner orients the sailplane toward the predicted thermal and attempts to detect when the plane is within the thermal. Then, the low-level planner latches onto the pocket of rising air and circles until the thermal dissipates.
In addition, because these systems adapt and learn as they gain experience, the sailplane’s AI actually becomes a better pilot with every flight.
“The system will perform better on Friday than on Thursday because it incorporates information based on past flights,” said Andrey Kolobov, one of the head researchers on Microsoft’s sailplane project.
In fact, the sailplane improved so much during the two days of testing that researcher and pilot Rick Rogahn remarked, “We’ve reached the point this summer where […] [t]he algorithm is doing better than me as a sailplane pilot.”
The success of this AI is so monumental because it surmounts a challenge with which computers have struggled wince their very inception: uncertainty.
Programmers find it relatively easy for AI’s to master games such as chess or go because the rules are clear-cut and all relevant information is already set before them. In contrast, the sailplane’s AI has incredibly limited information with which to work. For example, it may know precisely what direction the wind is blowing at its location, but what about a quarter of a mile to the east? To the south? What about 100 feet higher up?
In addition to dealing with uncertainty, the sailplane has to plan for the distant future in what is referred to as “sequential decision making under uncertainty.”
“It’s really the question of, ‘How do you plan for the future, several steps ahead?’” Kapoor pointed out. “Computationally, that’s a very hard problem.”
And yet, despite all of the challenges they faced, these researchers managed to triumph over that problem to build a sailplane that rivaled and nearly beat the world record for longest automated glider flight, which still stands at over five hours.
While the sailplane is undeniably an impressive feat of engineering, it is not game-changing in and of itself. Rather, the programming ideas that created it are the true breakthrough; they are paving the way to the nearly boundless future of AI.
“For us, the sailplane is a testbed for technologies at the core of anything that will be considered intelligent in the next 10 years,” said Andrey Kolobov, one of the head researchers of the sailplane project.
This unique ability to plan for and choose among uncertain outcomes is precisely what true artificial intelligence of the future will have do, and this glider is a monumental step toward that goal.