High-altitude platform stations (HAPS) are effective tools for communication and surveillance because they operate from the stratosphere, usually at around 12 miles (20 km) above the Earth, much closer than satellites. There are two types of HAPS: lighter-than-air (LTA) HAPS, such as high-altitude balloons, and heavier-than-air (HTA) HAPS, such as fixed-wing aircraft.
LTA-HAPS are generally solar-powered, unmanned balloons that use gases such as hydrogen or helium – hence, lighter than air – to maintain buoyancy. But, they are at the whim of the winds. The way LTA-HAPS maintain their position relies on a complex interaction between naturally occurring phenomena and AI.
HAPS combine the flexibility of aircraft with the endurance of satellites. For that reason, they can be used for overhead communications like telephone and internet services or radio station broadcasts, as well as for Earth observation and military intelligence-gathering, reconnaissance and surveillance (IRS).
In order to deliver internet connectivity from the stratosphere to the world’s hard-to-reach places, Loon, a subsidiary of Alphabet Inc., spent nine years in collaboration with Google Brain developing the LTA-HAPS technology that would achieve this.
There are obvious challenges to relying on a solar-powered, propellor-less, unmanned balloon to provide these types of services; the primary one is the weather, which cannot be controlled. The balloon needs to be able to withstand high winds and huge swings in temperature over the long term while delivering constant connectivity to the ground below. Relying solely on weather forecasts can be dangerous, as they are often wrong.
Additionally, the balloon does not have an infinite onboard energy source. It depends on the sun to provide the power for both navigation and communication, so to have the balloon correct itself if it is blown off course is a waste of precious power.
Research scientist Marlos Machado was part of the team that developed the technology to ensure that the balloon did what it was designed to do, with the greatest efficiency.
“We want the balloon to be in a specific position,” Machado said. “There is a catch, though. The problem with these balloons is that they don’t have propellors. The only way that they [can] navigate in the stratosphere is by sailing the winds.”
By "sailing the winds", Machado is referring to the way the balloon harnesses the wind’s strength, altitude, and direction to either maintain the desired location – called station-keeping – or travel to a new one.
These balloons are huge, the size of a tennis court. In the stratosphere, winds are blowing in different directions at different altitudes and speeds and, in terms of controlling its own movement, the balloon is limited to two directions: up and down.
“The notion is that you have some lighter-than-air gas inside the balloon, meaning that if you just leave the balloon by itself, the balloon [will] go up,” Machado said. “If you want the balloon to go down – to sink – all you have to do is to pump air into the balloon. It has a fixed volume, meaning that it’s heavier for the same volume.”
Introducing a lighter-than-air gas into a bladder inside the balloon causes it to rise into the wind stream required to move it to the correct position. If the balloon needs to go down so that it can be carried by the wind in a different direction, ambient air is pumped into a fixed envelope that sits inside the base of the balloon to provide ballast, or a valve opens on the LTA bladder, expelling some of the gas so that the balloon becomes heavier.
To gain some control over this somewhat limited maneuverability, the researchers developed a reinforcement learning AI that rewarded the balloon’s behaviors. A simulator combined weather forecasts with observed weather data and a Gaussian process to provide the best wind predictor. If the balloon responded "correctly", it was rewarded. On occasion, however, the wind wins.
“Sometimes there is nothing you can do,” Machado said. “If the winds are not in your favor, there is nothing you can do but wait until good winds show up.”
Ultimately, the research was a success. In 2017, the team navigated balloons to provide emergency connectivity to hundreds of thousands of people following flooding in Peru and a major hurricane in Puerto Rico.
Unfortunately, years of development proved to be a costly exercise, and in 2021, Loon discontinued work on their LTA-HAPS after the company was shut down. Other companies have picked up the mantle, continuing to develop their own LTA-HAPS for use in emergency communications, disaster recovery, providing private wireless networks, and extended offshore coverage.
The study by Machado and colleagues was published in the journal Nature, and in the video below Machado provides a more in-depth explanation of his research on developing an autonomous navigation system using reinforcement learning. The lecture is part of the AI Seminar Series delivered by Amii Intelligence.
Source: Nature
