Designing Vision-Based Navigation Policies for Resource-Constrained Underwater Robots

Abstract

Coral-reef monitoring requires autonomous platforms capable of sustained, fine-grained data collection in environments where human divers cannot persist. Bio-inspired underwater robots such CoralBot; a fish-shaped platform with fin-based actuation, dual fisheye cameras, and onboard Raspberry Pi5, offer a compelling path forward, but enabling autonomous navigation on such resource constrained hardware requires solving perception, control, and validation challenges simultaneously. This thesis presents an end-to-end vision-based navigation pipeline for CoralBot, beginning with a systematic evaluation of the Depth-Anything-V2 monocular depth foundation model on underwater fisheye imagery, including input size studies and calibration analysis. The depth pipeline feeds a six-region detection system that triggers heuristic obstacle avoidance, validated through physical pool deployment on CoralBot at two venues. To extend behavior beyond reactive avoidance, we develop a hardware-matched HoloOcean simulation in which the constrained HoveringAUV agent serves as a CoralBot proxy and train a twelve-model behavioral cloning sweep across input modality (depth versus RGB), lighting augmentation, backbone initialization, and training budget. Closed-loop evaluation across seen and held-out trajectories produces two ImageNet-pretrained depth policies that achieve 10/10 and 9/10 trajectory survival with collision rates 0.1% and 0.2% and identifies depth-channel mode collapse as the binding constraint on tasks requiring vertical motion. As a downstream application, we benchmark COLMAP-based 3D reconstruction across terrestrial, in-air, and underwater datasets, characterize the failure modes that prevent recognizable underwater reconstructions, and propose a SIFT-based feasibility diagnostic. The thesis concludes with a proposed deployment architecture for physical CoralBot and mitigations for the sim-to-real gap and the mode-collapse limitation.