Date of Award
Spring 2019
Project Type
Dissertation
Program or Major
Computer Science
Degree Name
Doctor of Philosophy
First Advisor
Wheeler Ruml
Second Advisor
Momotaz Begum
Third Advisor
Laura Dietz
Abstract
Autonomous robotic systems are becoming widespread in the form of self-driving cars, drones, and even as consumer appliances. These systems all have a planner that makes them autonomous. The planner defines the way these robots evaluate and select among the possible actions that are available to them.
This dissertations is about a specific type of planning called on-line real-time planning that is especially applicable to autonomous robots.
The central thesis of this work is that real-time heuristic search can be a viable planning method for complex state spaces. Planning for autonomous agents requires novel methods that are not direct adaptations of off-line planning methods but designed specifically for the task to provide fast and reliable execution while keeping the agent safe to guarantee that the goal will be reached. While there are many approaches to planning for embodied agents, my work pursues a class of on-line planning methods called real-time heuristic search that provide real-time bounds on action selection time.
This dissertation makes three major contributions.
Traditional real-time search algorithms are not guaranteed to recognize a subgraph from which the goal is not reachable before entering it, thus they are inherently incomplete in domains with dead-ends when a time bound is imposed on the planner. The first contribution of my dissertation addresses this issue by introducing a real-time search method with completeness guarantees in domains with dead-ends.
Real-time planning in the presence of local minima is particularly challenging due to the bounded rationality of real-time decision making. The second contribution of my dissertations is to introduce novel methods to mitigate this deficiency of real-time search.
Real-time search methods should be designed specifically to optimize the metric of interest: goal achievement time. Having more time to think leads to more informed and better quality decisions that can improve the overall goal achievement time. The third contribution of my dissertation is the introduction of a real-time metareasoning technique that considers actions that do not lead to the best discovered node, according to the commonly used f metric, in order to provide more time to the agent to plan the upcoming iteration.
Together these contributions support that: real-time heuristic search is a viable planning method for complex state spaces.
Recommended Citation
Cserna, Bence, "Real-time Planning For Robots" (2019). Doctoral Dissertations. 2442.
https://scholars.unh.edu/dissertation/2442