Navigating stratospheric balloons with reinforcement learning

Superpressure balloons can operate autonomously in the stratosphere for months, making them a cost-effective platform for communication, Earth observation and gathering meteorological data. In a Loon superpressure balloon, vertical motion is achieved by pumping air in and out of a chamber, whereas horizontal motion is dictated by the wind, and so flight controllers must ascend and descend to find and follow favourable currents for navigation. Despite these simple dynamics, long-term balloon control is extremely challenging, particularly the task of station-keeping. Marc Bellemare and colleagues now demonstrate a high-performance flight controller that uses reinforcement learning, which they deploy to station Loon balloons at multiple locations across the globe, including a 39-day controlled experiment over the Pacific Ocean, with data augmentation and a self-correcting design used to overcome the challenge of learning from imperfect data.