Date of Project

3-28-2025

Document Type

Honors Thesis

School Name

College of Arts and Sciences

Department

Computer Science

Major Advisor

Dr. Sayani Sarkar

Second Advisor

Dr. Nathan P. Johnson

Abstract

Autonomous Underwater Vehicles (AUVs) face significant challenges in underwater navigation, including generating smooth paths, avoiding obstacles, and adapting to complex conditions. This paper introduces a hybrid path-planning algorithm, D-RL*, that integrates the D* Lite algorithm for efficient initial pathfinding with Deep Reinforcement Learning methods to refine paths for smoother trajectories. The proposed approach addresses D* Lite's inability to produce continuous, smooth paths and baseline Reinforcement Learnings’ failures in environments requiring significant detours. Experimental results in four progressively complex environments highlight D-RL*’s ability to plan smoother paths than D* Lite while training in a shorter amount of time and generating shorter paths than baseline Deep Reinforcement Learning methods. These findings demonstrate D-RL*’s potential to deliver optimal and reliable navigation, making it well-suited for applications such as environmental monitoring and disaster response. Future work will extend testing to dynamic environments, three-dimensional navigation, and kinematic constraints to further enhance its operational feasibility.

Share

COinS