Learning to Soar: Exploration-Exploitation Algorithms for Autonomous Soaring Flight

Опубликовано 12 июля 2016, 22:50

Soaring is the process of collecting energy from the wind during flight with an aerial platform. Animal studies have demonstrated that large birds can significantly extend their flight duration by soaring in favourable wind conditions. Since most UAVs have their flight duration limited by some form of internal energy storage for propulsion (batteries or fuel), autonomous soaring provides a unique opportunity to extend flight duration without modifying payload. This talk will present work performed at the University of Sydney towards understanding soaring and developing autonomous soaring algorithms. We will present some basic reactive strategies for dynamic soaring and attempt to approach the more complex problem of simultaneously exploring (mapping) and exploiting (soaring) in unknown wind fields. We will explore Gaussian Process-based mapping along with simple mode-switching and more advanced reinforcement learning soaring approaches. We will also discuss higher-level mission planning approaches using Monte Carlo Tree Search for thermal soaring exploration problems.