Unsupervised Online System Identification for Underwater Robotic Vehicles

Abstract

This paper presents an unsupervised strategy for online system identification of underwater robotic vehicles. The proposed method consists of four operating modules, namely the state estimation module, the collision avoidance module, the excitation module, and the parameter estimation module, that collaborate online in a closed-loop architecture. The excitation inputs are formulated online, based on the convergence achieved for each actuated degree of freedom, thus speeding up the identification process while concurrently evading overfitting. Additionally, the proposed algorithm guarantees safe operation by avoiding collisions with the workspace boundaries; hence, no human supervision is required during the identification procedure. Moreover, the overall scheme is of low complexity and can be easily integrated into the real-time embedded system framework of underwater robotic vehicles. Finally, the efficacy of the proposed strategy was experimentally verified via an online system identification of a small underwater robotic vehicle, and the accuracy of the estimated parameters was further experimentally evaluated via a trajectory tracking task, by comparing the performance of a PID control scheme that employed feedforward compensation of the identified dynamics, with the corresponding performance of a conventional model-free PID control scheme.