Online Deep Learning Control of an Autonomous Surface Vehicle Using Learned Dynamics

Realtime model learning is a challenging task for autonomous surface vehicles (ASVs) sailing in a variable sea environment. Deep learning based on deep neural networks (DNNs) takes advantage of high representative capabilities. However, it is difficult to achieve stable learning control performance due to modeling errors or model bias. On the other hand, extended state observer (ESO) takes advantage of fast reconstructing unknown disturbances. In this paper, an online deep learning control method is presented for an ASV to achieve trajectory tracking. Specifically, a general DNN is constructed at first to learn the unknown ASV dynamics online with the collected data one by one at each time to improve scalability. Then, an ESO is designed to estimate the modeling errors of the DNN for improving the model learning accuracy further. Finally, a stable online deep learning trajectory tracking control law is designed based on the learned ASV dynamics from the DNN and the reconstructed modeling errors from the ESO. By using the cascade theory, it is proven that the closed-loop trajectory tracking control system is input-to-state stable and all signals are uniformly ultimately bounded. Simulation results of the circular trajectory tracking show that the proposed method improves the transient tracking performance compared with the DNN-based and ESO-based control methods. Moreover, an "8-type" trajectory tracking simulation is further provided to demonstrate the generalization capabilities of the proposed method for new trajectories and new environments.