Committee Chair: Katherine Skinner

Abstract:
Underwater robots perform crucial search, inspection, and manipulation tasks in environments that are dangerous, impractical, or impossible for humans to operate in. In order to safely perform useful work, it is crucial for these robotic systems to be equipped with algorithms that enable spatial awareness, decision making, and localization. Alternate sensing modalities such as sonar address some of these challenges but introduce new trade-offs: acoustic noise, view-dependent imaging, and lack of large datasets for machine learning methods.

This dissertation addresses challenges associated with autonomous survey, reacquisition, and inspection missions for underwater robots. Autonomous survey encompasses the deployment of robots to capture imagery of a large search area. Reacquisition refers to the maneuvers to capture varying views of any found objects. Finally, close-up inspection of objects involves capturing optical and acoustic imagery of objects for downstream tasks. These tasks are crucial to search and rescue, oceanography, defense, and infrastructure inspection applications. Traditionally, search for submerged objects of interest has been carried out with ship-mounted sonar and human divers. However, this manual process is time consuming, dangerous, and does not scale to large search areas. The use of Autonomous Underwater Vehicles (AUVs) promises to speed up the process of underwater survey and inspection. However, the data these vehicles collect often requires manual analysis by humans, which can take weeks and does not enable in-situ autonomous decision making. Although machine learning methods can speed up decision making and data processing, there are several challenges for deploying machine learning methods in the field. After an item has been found, marine robotic systems can be deployed for a close-up inspection to enable detailed surveys of targets of interest.

This work contributes several novel methods that aim to enable autonomous search and survey for marine robotic systems. The first contribution of this work is a method that addresses the lack of training datasets for learning-based object identification in sonar imagery collected from robotic search missions. Specifically, this work presents a method for shipwreck segmentation that is trained entirely in simulation and performs zero-shot shipwreck segmentation with no additional fine-tuning on real data. The second contribution is a diverse, open-source dataset and benchmark for shipwreck segmentation in sonar imagery. This contribution establishes an open-source benchmark and enables future research in machine learning for ocean exploration. After an object is found, operators may perform lower-altitude, specialized survey patterns for imaging the object of interest. The third contribution is an adaptive surveying algorithm that plans more efficient re-inspections of detected targets using side scan sonar. This algorithm is a novel active perception method for side scan sonar that both chooses the next best view to maximize classifier performance and aggregates and classifies multiple views from an adaptive survey. Finally, we develop a novel Gaussian splatting framework for close-range inspections using imaging sonar that demonstrates realistic novel view synthesis and models acoustic streaking phenomena. This method can enable synthetic data generation of realistic imaging sonar data for training deep learning models and additionally enables denoising of captured sonar data, allowing for increased interpretability. The methods presented throughout this thesis are evaluated on real-world datasets collected from field work experiments using marine robotics platforms. This dissertation represents progress toward fully autonomous underwater systems that can operate with minimal human oversight in order to greatly accelerate search and survey missions for critical underwater applications.

Zoom link: https://umich.zoom.us/j/97161696814