Reinforcement Learning for Flight Control

Abstract

In recent years Adaptive Critic Designs (ACDs) have been applied to adaptive flight control of uncertain, nonlinear systems. However, these algorithms often rely on representative models as they require an offline training stage. Therefore, they have limited applicability to a system for which no accurate system model is available, nor readily identifiable. Inspired by recent work on Incremental Dual Heuristic Programming (IDHP), this paper derives and analyzes a Reinforcement Learning (RL) based framework for adaptive flight control of a CS-25 class fixed-wing aircraft. The proposed framework utilizes Artificial Neural Networks (ANNs) and includes an additional network structure to improve learning stability. The designed learning controller is implemented to control a high-fidelity, six-degree-of-freedom simulation of the Cessna 550 Citation II PH-LAB research aircraft. It is demonstrated that the proposed framework is able to learn a near-optimal control policy online without a priori knowledge of the system dynamics nor an offline training phase. Furthermore, it is able to generalize and operate the aircraft in not previously encountered flight regimes as well as identify and adapt to unforeseen changes to the aircraft’s dynamics.