Loon’s mission is connecting people everywhere using audacious new technology. We use a synthesis of machine learning and automation with a superpressure balloon-based aircraft to create a unique high altitude pseudo-satellite (HAPS) that can provide connectivity, earth observation, collect weather data to improve forecasts, and perform other tasks from the stratosphere. This technology would not be possible without software automation, and as a result research-level machine learning has a great deal of applicability at Loon. A prime example of this is our Nature publication, released earlier this week, that describes how we have used deep reinforcement learning to improve station keeping for our HAPS system in real flight through the stratosphere and across our production fleet. In this talk I’ll discuss some of the areas and applications where machine learning can further improve Loon, and dive into the technology described in the Nature paper as an example of successful collaboration between research and the core Loon technology stack that led to deep reinforcement learning taking flight with Loon.