Implementing Control-oriented Meta-learning on Hardware

Abstract

Unpredictable weather conditions pose a daunting challenge for the robust control of unmanned aerial vehicles, also known as drones. The control-oriented meta-learning algorithm aims to solve this problem by learning a controller that can adapt to dynamic environments. This algorithm has already been derived and simulated for a two-dimensional model. This project explores the implementation of the control-oriented meta-learning algorithm on a hardware platform. After extending the algorithm to a three-dimensional model, it was tested in a physics-based simulator and deployed on a hexarotor in the real world. Both in simulation and in real life, the learned controller outperformed a traditional controller in the presence of wind.