Embodied Hydrodynamic Reservoir Computing for Underwater Obstacle Localization

Abstract

In underwater environments, fluid–structure interactions can provide detailed information about the surroundings. However, exploiting these hydrodynamic responses as a contactless sensing mechanism for obstacle localization remains challenging. Existing approaches often require complicated sensors or data processing methods to operate reliably in underwater environments. This study presents the design and performance evaluation of three embodied soft panels, inspired by the physiology and sensory setup of a fish's body. We demonstrate that embodied intelligence can be achieved by using a framework called physical reservoir computing (PRC), which utilizes the physical body dynamics as a computational resource. Herein, the 3D-printed panels with embedded sensing networks were designed for the pur pose of extracting their body-state information for embodied computation. The nonlinear autoregressive moving average (NARMA) task was used to evaluate their computational per formance. We show that under appropriate physical reservoir dynamics, the panels exhibited greater capability to emulate nonlinear dynamical systems and function as physical reservoir computers. Furthermore, we also show that the panels had enough computational power to estimate the position of a nearby obstacle based on variations in their body dynamics. Our results suggest that the panels can serve as potential frameworks for intelligent swimmers capable of perceiving their environments through the self-sensing mechanism and estimating relevant information.