03/18/2026
By Nolan Allen

The Kennedy College of Science, Miner School of Computer and Information Sciences invites you to attend a Master's Thesis defense by Nolan Allen on “BLUE: Behavior Learning for Underwater Execution."

Candidate Name: Nolan Allen
Degree: Master’s
Defense Date: Monday, March 30, 2026
Time: 10 a.m. ET
Location: Dandeneau 309 and via Zoom
Thesis/Dissertation Title: BLUE: Behavior Learning for Underwater Execution

Committee:

  • Advisor: Reza Azadeh, Ph.D., Miner School of Computer and Information Sciences, University of Massachusetts Lowell
  • Maru Cabrera, Ph.D., Miner School of Computer and Information Sciences, University of Massachusetts Lowell
  • Paul Robinette, Ph.D., Department of Electrical and Computer Engineering, University of Massachusetts Lowell

Abstract:
Autonomous underwater manipulation is a challenging domain and an open problem in research, with performance often degraded by factors such as unpredictable underwater disturbances, currents, visual distortions, and sensing limitations. To this end, we propose a robust and modular underwater manipulation framework. Existing approaches often require extensive tuning or training data, use specific hardware setups, are unequipped to handle unexpected base movement and drift, or are unable to solve and generalize complex skills in novel task configurations. The inclusion of these design constraints sets our framework apart as a novel approach to address the challenges presented by practical deployment in underwater environments.

To achieve robust skill encoding and generalization, our framework integrates high-level STRIPS-based symbolic planning to generate a sequence of predefined movements, a Learning from Demonstration (LfD) model to learn basic task movements from human demonstration and generalize them to new task configurations, and an optimal control module using a Time-Varying Linear Quadratic Regulator (TVLQR) to recover smoothly from deviations in the planned trajectory while completing the movements. We further mitigate sensor noise with an Unscented Kalman Filter (UKF) to improve task pose estimation. To demonstrate our framework as learning model-agnostic, we implement three LfD models—specifically, Laplacian Trajectory Editing (LTE), Dynamic Movement Primitives (DMP), and Jerk-Accuracy (JA)—each with a unique approach. This unified and synergistic framework enables robust and efficient training and autonomous completion of a variety of underwater manipulation tasks.

We validate the framework on two industry-relevant tasks, valve turning and loop-on-hook placement, using a Reach Alpha 5 manipulator mounted on a BlueROV underwater platform. Experimental results in laboratory and pool environments confirm the robustness, data-efficiency, adaptability, and practical potential of the framework to advance autonomous underwater manipulation. These results demonstrate our work as a useful contribution to the reliability and autonomy of underwater robotic systems.