Loon is a company that’s been working on a project for a while aiming to use superpressure balloons floating in the stratosphere to deliver Internet connectivity. We don’t hear about the project very often, and some may wonder if it is still relevant with SpaceX already in trials with its Starlink Internet service using satellites. Loon has deployed something very interesting with what the company claims is the world’s first deployment of a reinforcement learning aerospace system.
The company has long had a goal of efficiently steering its balloons to locations needing connectivity. The challenge is that navigating a super pressure balloon through the stratosphere has two options: up or down. Navigation is complex, and the task is well-suited to automation. Loon balloons follow the prevailing wind.
Loon says that a small group within its company and Google AI have been working together to develop a more powerful navigation system leveraging deep reinforcement learning (RL). RL is a type of machine learning that enables an agent to learn by trial and error in an interactive environment via feedback from its own actions and experiences. The company admits that it wasn’t clear that deep RL was practical or viable for its high-altitude balloons early on.
Testing has proven that RL is practical for a fleet of stratospheric balloons. The navigation system in use today faces a complicated task solved by an algorithm that learns via a computer experimenting with balloon navigation in a simulation. The task in front of the RL system is highly complicated as the balloon often lacks the power required for the ideal maneuver.
That means that frequent decision opportunities span a long planning horizon. Despite the complexities, reinforcement learning has promised to help Loon steer balloons more efficiently than human-design algorithms in use widely now. RL can quickly enable the navigation system to manage new aircraft and manage different uses of the Loon fleet. So far, the RL controller has navigated a group of balloons for nearly 3000 flight hours with what the company calls excellent results.