Learning to soar

The strategy used by soaring birds to find and navigate thermals within complex natural landscapes of convective currents is not yet fully understood, and is thus difficult to ‘teach’ to man-made vehicles successfully. In this context, approaches based on reinforcement learning can help to identify effective navigational strategies in the form of sequences of decisions informed by specific environmental cues. Here, Massimo Vergassola and collaborators present a reinforcement learning approach that enables a suitably equipped glider to navigate thermals in an open field autonomously. For this purpose, the authors developed on-board methods for estimating local vertical wind accelerations and roll-wise torques on the glider, which are shown to act as effective navigational cues. This type of navigational strategy is likely to support further research into autonomous soaring vehicles.