Prediction of Hydrodynamic Loads on a Flexible Bio-Inspired Underwater Propulsor Using Physical Reservoir Computing

Abstract

Abstract Bio-inspired autonomous underwater vehicles (BAUVs) have the potential for stealthy navigation and operation in sensitive underwater environments by mimicking the natural behaviors and compliance of marine animals. However, BAUVs still face challenges where maneuverability and control are difficult to achieve, especially in high-energy coastal zones. Maneuverability and control could be improved by implementing real-time sensing of the hydrodynamic forces acting on a BAUV’s compliant body. In this study, we predict the hydrodynamic forces generated by a soft-robotic propulsor using Physical Reservoir Computing (PRC), with the future aim of improving the propulsor’s maneuverability. Physical reservoir computing is a machine learning technique that utilizes the nonlinear dynamics of a physical system for information processing. Compared to software-based machine learning frameworks, PRC requires significantly lower computational costs. With PRC, a majority of the information is processed within the physical system itself, enabling real-time prediction of time-varying target signals. In this project, we use PRC to leverage the nonlinear fluid-structure dynamics of a soft-robotic propulsor to predict its time-varying thrust. Our soft-robotic propulsor is powered by an antagonistic pair of embedded hydraulically amplified self-healing electrostatic (HASEL) actuators. We conducted experiments to measure the thrust generated by the soft robot in quiescent water. Using the PRC framework, we predicted the thrust using a linear weighted sum of the time-varying kinematic strains measured at multiple locations on the propulsor’s compliant body. We trained the linear weights by minimizing the error between the PRC prediction and the experimentally measured thrust. We then analyzed various sensor placement configurations to achieve the best performance. By using the PRC framework, we were able to accurately predict the time-varying thrust. Using the sensor configuration that yielded the best performance, we achieved R-squared errors of 0.908 and 0.898 on the training set and test set, respectively. This PRC method holds significant potential for predicting time-varying hydrodynamic loads in real-time which can be used to improve control and maneuverability of BAUVs.