Quadcopter disturbance estimation using different learning methods

Abstract

Precise modeling of quadcopter dynamics is challenging due to the complex nature of its construction, aerodynamic effects, friction at rotors, and wind effects involved. In general analysis, these unmodeled dynamics are kept as external disturbances to the system. Machine learning techniques can effectively be used to estimate or predict the unknown kinetic effects in the quadcopter dynamical model. This paper attempts to compare the effectiveness of two popular machine learning techniques in modeling vehicle dynamics, namely neural networks (NN) and Gaussian process regression (GPR). The dynamic model of the quadcopter is expressed as a combination of a known nominal model and the unknown term, which was learned separately using the two methods. The performance of these two approaches is evaluated using a dataset collected by manually flying the AscTec Hummingbird quadcopter under an OptiTrack motion capture system. The learning process has been performed off-line, and a performance comparison between NN and GPR is discussed in the paper.