Self-Organizing Swarms & Robotics Lab
Designing Vision-Based Navigation Policies for Resource-Constrained Underwater Robots
Senior Thesis · Advised by Prof. Radhika Nagpal
Coral-reef monitoring needs autonomous platforms that can collect fine-grained data where human divers cannot persist. CoralBot — a fish-shaped robot with fin-based actuation, dual fisheye cameras, and an onboard Raspberry Pi 5 — is the hardware this thesis builds for. Getting it to navigate on its own means solving perception, control, and validation together.
The pipeline starts with a systematic evaluation of the Depth Anything V2 monocular depth model on underwater fisheye imagery. That depth signal feeds a six-region detection system that triggers obstacle avoidance, validated through physical pool deployments at two venues.
To go beyond reactive avoidance, I built a hardware-matched HoloOcean simulation and trained a twelve-model behavioral cloning sweep. Two ImageNet-pretrained depth policies reached 10/10 and 9/10 trajectory survival at 0.1% and 0.2% collision rates — with depth-channel mode collapse emerging as the binding constraint on vertical motion. A final COLMAP reconstruction benchmark characterizes why underwater scenes fail and proposes a SIFT-based feasibility diagnostic.