High-flying balloons are bringing broadband connectivity to remote nations and post-disaster zones where cell towers have been knocked out. These “super-pressure” helium-filled polyethylene bags float 65,000 feet up in the stratosphere, above commercial planes, hurricanes, and pretty much anything else. But keeping a fleet of tennis-court-sized, internet-blasting balloons hovering over one spot has been a tricky engineering problem, just like keeping a boat floating in one place on a fast-moving river.
Now researchers at Google spinoff Loon have figured out how to use a form of artificial intelligence to allow the balloon’s onboard controller to predict wind speed and direction at various heights, then use that information to raise and lower the balloon accordingly. The new AI-powered navigation system opens the possibility of using stationary balloons to monitor animal migrations, the effects of climate change, or illegal cross-border wildlife or human trafficking from a relatively inexpensive platform for months at a time.
“It’s super hard to have the [balloon] network over the people who need connection to the internet and not drifting far away,” says Sal Candido, chief technology officer at Loon. The high-tech balloons were tested last year over Peru and managed to stay on target without a human controller. Because winds blow in various directions at each altitude, the AI-based controller was programmed to use reinforcement learning, or RL, to search a database of historic records and current weather reports to predict the best elevation to keep the balloon in one place. It also checked how much electricity the balloon’s solar panels were generating to operate the device’s instruments.
“What the RL is doing for us is deciding what’s the situation with the balloon, how much power does it have left, what is the best action that the balloon could do right now to stay over the person with the cellphone in their hand,” says Candido about how Loon keeps the balloons over broadband customers.
Candido is coauthor of a journal article on the computer programming experiment published today in the journal Nature. The study details a 39-day experiment over the Pacific Ocean in which an AI-based Loon balloon was parked over a spot along the equator and received information from other balloons in the area. The balloon was able to stay close to its target by performing a series of figure eights as it moved up and down the atmosphere. Since the AI agent didn’t have a complete record of wind direction and speed in the remote area, it filled in the gaps by adding randomly generated “noise” to the current wind data, to better map out the range of wind speeds and directions that could plausibly occur and to improve assessments of the variety of paths the balloon might take in the future. The algorithm improved decisionmaking time during flights, compared with Loon’s previous balloon navigation systems, which did not use reinforcement learning.
Candido and his team have been working on this problem for several years, since the company was first launched as part of the Google X research lab in 2012. Loon is now a subsidiary of Google’s parent company Alphabet.
The big advancement since then has been applying reinforcement learning, something previously used in video games, to a real-world challenge, according to Marc Bellemare, lead author on the Nature paper and a research scientist at Google Canada. “Machine learning refers to the idea of taking data and making predictions about outcomes,” Bellemare says. “With reinforcement learning we are focusing on the decision part. How do we go up or down based on that data? Not only is [the AI controller] making decisions, but making decisions over time.”
Some experts believe the AI-powered balloons can also be used to monitor Earth’s environmental vital signs, such as checking on melting in the Arctic permafrost, the exchange of greenhouse gases from tropical rainforests, or even the atmospheric pressure and wind currents that give rise to powerful hurricanes in the Atlantic and Pacific. The advantage of this kind of new AI-based navigation system is that the balloons can be deployed from a launchpad far away—such as Puerto Rico or Nevada, where Loon operates—and then actively surf winds to reach their target, much like how a sailboat tacks against the wind to cross the ocean.
“You can launch them where it’s convenient and cheaper, and then they could move themselves,” says Scott Osprey, a climate scientist at the University of Oxford, who was not involved in the research with Loon. Osprey sees a big role for stationary balloons in recording seismic waves from active volcanoes, for example, or even on interplanetary missions to explore cloudy atmospheres. “You could send a probe to Venus and look above the cloud tops in orbit,” he says, “or station it there for months at a time and communicate with a satellite above.”
A trip to Venus may be a few years down the road. Today, however, the new AI-powered autonomous balloons are stationed above Kenya, providing internet service to customers of Kenya Telecom. The firm recently broke a record for keeping one of its balloons aloft for 312 days and is expanding service to nearby Mozambique in the coming months.
Google’s Bellemare says the advent of using reinforcement learning to navigate the balloons for long periods of time will open up all kinds of applications for scientific remote sensing and commercial projects as well. For him, it’s another step in making truly intelligent machines to perform difficult tasks without a human controller behind them. “The really exciting thing is using reinforcement learning,” he says. “It’s like if you are trying to learn to ride a bicycle—it’s harder to write down the equations. It’s easier just to try it out.”