Date of Award
2025
Document Type
Dissertation
Degree Name
Doctor of Philosophy (PhD)
Department
Computer Science
Committee Chair
Vineetha Menon
Committee Member
Farbod Fahimi
Committee Member
Jacob Hauenstein
Committee Member
Chaity Mukherjee
Committee Member
Howard Chen
Research Advisor
Vineetha Menon
Subject(s)
Robots--Motion, Deep learning (Machine learning), Reinforcement learning
Abstract
Navigation methods are needed for robots to operate in a physical world. Path planning for navigation can be a difficult problem to solve due to uncertainties in the world and the robot. Navigation also must be completed in real time to ensure the robot does not collide with other entities in the world. Declarative re-planning (DRP) is a method for path planning. It leverages hierarchical temporal abstractions to drive learning, and run time efficiencies. It is multiple orders of magnitude more sample efficient compared to non-temporally abstract methods, and the asymptotic run time efficiency is linear for the full navigation process. The design of reward functions are also investigated. The trade off between dense reward functions being easier to learn but may drive side effects and sparse reward functions that hard to learn but are more robust to side effects. DRP drastically reduces the penalty for using sparse reward functions enabling the agent designer to leverage the advantages of sparse reward functions with out the cost. The hierarchical temporal abstraction also yields a heuristic for selecting an optimal discount factor for sparse reward functions. This work demonstrates how DRP is effective at zero shot transfer learning. DRP is able to train a navigation policy on one robot, and be used on other robots iii while remaining equally effective at reaching its goal. The other robots do not need to have the same properties, actuators, or number of actuators. DRP also enables simulation acceleration by allowing lower fidelity trajectory calculations to be used without effecting the outcome of the navigation trajectory. Finally, how navigation can be made trustworthy via the amount of information communicated and the nature of the information shared was investigated. More information improved responders ability to predict a robots future motion. No correlation was found between confidence and ability to predict the future motion.
Recommended Citation
Hemming, Nathan, "Declarative re-planning : a trustworthy deep reinforcement learning method enabling zero shot learning in mobile robot path planning" (2025). Dissertations. 472.
https://louis.uah.edu/uah-dissertations/472