Event

This dissertation explores the intersection of machine learning, perception, and simulation-based techniques to improve the efficiency of underwater robotics, with a focus on oceanic tasks. I begin by deploying marine object detection models on the BlueROV platform, using aerial imagery. The research then transitions to oyster detection, leveraging Oysternet and simulated data with Generative Adversarial Networks to enhance sim-to-real transfer and detection accuracy.

Next, I present an oyster detection system utilizing diffusion-enhanced synthetic data on the Aqua2 biomimetic robot for real-time, on-edge underwater detection. This system supports autonomous exploration, guided by an imitation learning framework that enables efficient navigation over oyster and rock reefs without relying on localization, improving information gathering while avoiding obstacles. A deep learning model for real/virtual image segmentation is also introduced to address water surface reflections, ensuring safe navigation in shallow environments.

Beyond marine applications, I apply these techniques to olive detection for yield estimation and industrial object counting, using simulated imagery. I also explore unresolved challenges like RGB/sonar data integration, proposing future research directions to further enhance underwater robotic learning through digital simulation. Through these studies, I demonstrate how machine learning and digital simulations can synergize to solve key challenges in underwater robotic tasks, advancing autonomous systems' ability to monitor and preserve marine ecosystems.