Skip to main content

This is how Google’s internet-serving Loon balloons can float for nearly a year

Only Google could think that the way to improve the flight of giant, helium-filled balloons is by coming up with better algorithms. And to be fair to the Mountain View-based search leviathan, it seems to have worked.

For the past couple of years, Project Loon, a subsidiary of Google’s parent company Alphabet, has been working to provide internet access in rural and remote parts of the world by using high-altitude balloons in the stratosphere to create aerial wireless networks. Last year, Loon announced that it had reached 1 million hours of stratospheric flight with its combined balloon fleet. Then, at the end of October, Loon set a new record for longest stratospheric flight by remaining airborne for a whopping 312 days, covering a distance of some 135,000 miles.

In a new article, published in the journal Nature, Loon explains just how its balloons are able to stay in the air for weeks at a time — without human intervention or full knowledge of surrounding winds. The secret? Some impressively cutting-edge A.I.

Catching currents

“Loon balloons navigate by moving up or down in altitude to catch favorable wind currents that take them in a desired direction,” Sal Candido, Loon’s chief technology officer, told Digital Trends. “The decisions about when to ascend or descend are determined by sophisticated algorithms. Traditionally, these algorithms have been written by human engineers. With reinforcement learning, we are leveraging A.I. to build these algorithms. In essence, we have built a machine that is capable of building a better navigation system than we humans can. That machine can also build these navigation systems in a fraction of the time that it takes us humans.”


Reinforcement learning is a flavor of machine learning heavily inspired by behaviorist psychology. Reinforcement learning’s guiding principle is the idea that software agents can learn to take action based around the maximizing of a reward. Famously, reinforcement learning was used by Google DeepMind to teach an A.I. to play classic Atari video games — using no more information than just the pixels that made up each frame of the games and the on-screen score. By being told to maximize its score, the DeepMind A.I. learned to play the games through trial-and-error, gradually honing its skills until it was a master.

Flying a balloon in such a way that it doesn’t get blown off-course is a far different task to playing computer games, of course. A successful balloon journey doesn’t come with a high score that makes it immediately apparent that it’s been successful. But, as Candido said, reinforcement learning is nonetheless a crucial part of Loon’s success.

“[Reinforcement learning] is able to process huge amounts of information and apply that to solving the problem, rather than a human needing to inherently understand how to react to that information or having a computer search the space of all possible outcomes,” he said. “Because Loon navigation improves by considering a huge number of factors and information [or] data, the complexity has surpassed what engineers are easily able to do [with regards to] the former, and the latter search is computationally difficult to scale across a full fleet. [That makes reinforcement learning] a great tool for the job.”

Making the right decisions

Using reinforcement learning, the artificially intelligent balloons are able to make optimal decisions about how to move, based on historical wind knowledge, observed and forecasted winds, and the projected future flight paths. All of this data is weighed up and different scenarios simulated before the balloon decides how to act.

Loon: 312 Days in the Stratosphere

Compared to the previous controllers used to control Loon, the new reinforcement learning-based methodology more effectively kept Loon’s balloons within range of their ground station so they could effectively send and receive signals. When they were knocked off-course, it additionally meant them returning faster to the right positions.

“Our new reinforcement learning-powered algorithm is active today, helping our balloons to stay above users in Kenya, whom we are serving as part of our partnership with Telkom Kenya,” Candido said.

Alphabet has long been committed to the idea of tech for good. The more people Loon can provide internet access to, the better the initiative will be. And, to do that, it needs ever-smarter technology driving it. As evidenced by this latest milestone, it seems to have all bases covered.

Editors' Recommendations