Title: Using Multi-Relational Embeddings as Knowledge Graph Representations for Robotics Applications

Committee:

Dr. Sonia Chernova (Advisor) - Georgia Institute of Technology, School of Interactive Computing

Dr. Zsolt Kira - Georgia Institute of Technology, School of Interactive Computing

Dr. Mohan Sridharan - School of Computer Science, University of Birmingham, UK

Dr. Matthew Gombolay - Georgia Institute of Technology, School of Interactive Computing

Dr. Devi Parikh - Georgia Institute of Technology, School of Interactive Computing

Abstract: User demonstrations of robot tasks in everyday environments, such as households, can be brittle due in part to the dynamic, diverse, and complex properties of those environments. Humans can find solutions in ambiguous or unfamiliar situations by using a wealth of common-sense knowledge about their domains to make informed generalizations. For example, likely locations for food in a novel household. Prior work has shown that robots can benefit from reasoning about this type of semantic knowledge, which can be modeled as a knowledge graph of interrelated facts that define whether a relationship exists between two entities. Semantic reasoning about domain knowledge using knowledge graph representations has improved the robustness and usability of end user robots by enabling more fault tolerant task execution. Knowledge graph representations define the underlying representation of facts, how facts are organized, and implement semantic reasoning by defining the possible computations over facts (e.g. association, fact-prediction). Existing knowledge graph representations used for robots are limited in their ability to scale to large knowledge graphs that model everyday domains, while accounting for fact uncertainty. This dissertation examines whether multi-relational embeddings can be used as representations of semantic domain knowledge to improve the task execution robustness of robots that must reason about fact uncertainty, scale reasoning to model everyday domains, adapt to changes in known facts, and be explainable to end users.