Autonomous underwater robots have received significant attention over the past two decades due to the increasing demand for environmental sustainability. Recently, gliding robotic fish has emerged as a promising mobile sensing platform in versatile aquatic environments. Such robots, inspired by underwater gliders and robotic fish, combine buoyancy-driven gliding and fin-actuated swimming to realize both energy-efficient locomotion and high maneuverability. In this dissertation, we first discuss the design improvements for the second-generation gliding robotic fish "Grace 2". These improvements have transformed the robots to underwater sensing platforms that can be modified to fit the requirements of a specific application with relative ease.We focus on the application of detecting and tracking live fish underwater, which is an important part of fishery research, as it helps scientists understands habitat use, migratory patterns, and spawning behavior of fishes. The gliding robotic fish has demonstrated its ability to detect special acoustic signals emulating tagged fish through a series of trials in Higgins Lake, Michigan. These tests have also validated a gliding-based strategy for navigating to a GPS waypoint, and offered insight into the limitations of the current design. Additional improvements are proposed to allow these robots to glide at larger depths and perform more interesting working patterns underwater.Motivated by the problem of tracking real fish, we consider the case where multiple robots localize and track a moving target without the need for a centralized node. We present theoretical treatment on how a network of robots can infer the location of an emitter (or target), and then track it, through a time-difference-of-arrival (TDOA) localization scheme in a fully distributed manner. In particular, we utilize a networked extended Kalman filter to estimate the target's location in a distributed manner, and establish that successful localization can be achieved under fixed and time-varying undirected communication topologies if every agent is part of a network with a minimum of 4 connected, non-coplanar agents. We further propose a movement control strategy based on the norm of the estimation covariance matrices, with a tuning parameter to balance the trade-off between estimation performance and the total distance traveled by the robots.Finally, motivated by the distributed localization problem, we investigate a more general problem of distributed estimation by a network of sensors. Specifically, we consider the class of consensus-based distributed linear filters (CBDLF) where each sensor updates its estimate in two steps: a consensus step dictated by a weighted and directed communication graph, followed by a local Luenberger filtering step. We show that the sub-optimal filtering gains that minimize an upper bound of a quadratic filtering cost are related to the convergence of a set of coupled Riccati equations. Then we show that the convergence of these coupled Riccati equations depends on the notion of squared detectability for the networked system, and proceed to provide necessary conditions that link the convergence of the coupled Riccati equations to the network topology and consensus weights of the communication graph.
Read
