Active Learning of Quadrotor Discrete Dynamics for Uncertainty-Aware Model Predictive Control

Alessandro Saviolo, Jonathan Frey, Abhishek Rathod, Moritz Diehl, Giuseppe Loianno

Abstract. Precise, safe, and agile control of quadrotors requires an accurate model of the system dynamics. In this paper, we propose a self-supervised framework to accurately model quadrotor's discrete dynamics purely from robot experience. The learned model is continuously refined during flight to compensate for model mismatches in presence of variations in the operating conditions. We design a model predictive control approach that incorporates the learned dynamics and is conditioned to their uncertainty for accurate closed-loop trajectory tracking. Therefore, the controller fully exploits the active learning strategy to boost its control actions employed to concurrently improve the model dynamics in a model-based reinforcement learning fashion. Experimental results demonstrate the feasibility of the proposed approach that accurately extracts the structure of quadrotor's dynamics from data while continuously adapting to environment disturbances and different flight regimes. This ensures accurate, safe, and agile flight by capturing effects that remain hidden to classical approaches.