Real-time dynamic model learning and adaptation for underwater vehicles

Abstract

Precision control of unmanned underwater vehicles (UUVs) requires accurate knowledge of the dynamic characteristics of the vehicles. However, developing such models are time and resource intensive. The problem is further exacerbated by the sensitivity of the dynamic model to vehicle configuration. This is particularly true for hovering-class UUVs since sensor payloads are often mounted outside the vehicle body. Methods are investigated in this thesis to learn the dynamic model for such a hovering-class UUV in real time from motion and position measurements. Several system identification techniques, including gradient estimation, Bayesian estimation, neural network estimation, and recursive linear least square estimation, are employed to estimate equations of motion coefficients. Experimental values are obtained for the surge, sway, heave, and yaw degrees of freedom. Theoretical results are obtained for the roll and pitch degrees of freedom. The experimentally obtained model is then compared to the true vehicle behavior.