Google AI Launches Pilot Program for Loon's Internet-Beaming Balloons

Alphabet's Loon has transitioned to an advanced navigation system for its internet-beaming balloons, moving away from human-designed algorithms to a cutting-edge artificial intelligence system developed in collaboration with Google AI. This new system employs a reinforcement learning (RL) model, now steering a fleet of balloons over Kenya, where Loon recently launched its first commercial service.

This marks a pioneering application of RL in a "production aerospace system," showcasing the potential of reinforcement learning in real-world scenarios. Previous RL systems have excelled in complex games like Go and Dota 2, and now Loon's AI optimizes balloon routes with remarkable speed and efficiency, enabling flights to cover similar or even greater distances while consuming less power. Notably, Loon recently set a record with a flight lasting 312 days.

Loon and Google AI utilized simulations to train the RL model through trial and error before initiating real-world tests in Peru. They compared its performance against a human-designed system known as StationSeeker during a 39-day evaluation over the Pacific Ocean. The AI outperformed StationSeeker by maintaining balloons within targeted areas for extended periods while minimizing energy consumption, crucial for delivering reliable internet coverage.

Unlike StationSeeker, which typically aimed directly for specific points and often overshot its targets, the AI focused on passive navigation. This approach conserved energy for critical moments and involved sophisticated maneuvers previously unknown to the Loon team, setting the stage for enhanced connectivity solutions.