“Underwater Localization and Mapping for Cost-Effective Robots” by A. Hinduja. Ph.D. dissertation, Carnegie Mellon University, Aug. 2024.
Autonomous Underwater Vehicles (AUVs) have become integral tools to solve real world tasks of surveying, mapping and inspection of human made infrastructure, as well as monitoring in natural environments. AUVs are often used in areas considered dangerous for human intervention, making the need for robust methods of perception and navigation paramount. Depending on the task, AUVs vary in size and form-factor which in turn also affects the onboard sensors used. For most infrastructure inspection tasks the most common types of sensors include Doppler Velocity Logs (DVL) and Inertial Measurement Units (IMU) for inertial navigation and acoustic sonars for perception. While these sensors represent the best options for AUVs, they have shortcomings such as often a prohibitively high cost, and a lack of information rich perceptual data when considering imaging sonars. We discuss the challenges faced in underwater localization and mapping with respect to imaging sonars, such as the lack of a general framework which works across different types of sonar makes, and also obstacles to engaging in research due to high setup and operational costs for AUVs. In this dissertation, we explore localization and mapping techniques designed for cost-effective underwater robots. We begin with looking at the problem of performing accurate Simultaneous Localization and Mapping (SLAM) when using sonar data due to featureless environments causing degenerate situations. I present a factor graph based solution to this problem. We then look at a framework for onboard acoustic localization of multiple ultra-low-cost underwater robots using off the shelf equipment. For improving feature-based SLAM using imaging sonars, I present a pose-supervised network to learn sonar image correspondences called SONIC. Lastly, this work is further expanded in C-SONIC to allow cross-sonar image correspondence. This allows cheaper robots with low frequency imaging sonars to find feature correspondences in maps made with high frequency imaging sonars.
BibTeX entry:
@phdthesis{Hinduja24thesis,
author = {A. Hinduja},
title = {Underwater Localization and Mapping for Cost-Effective Robots},
school = {Carnegie Mellon University},
type = {{Ph.D.}},
month = aug,
year = {2024}
}