As the scope of robotic applications in our world continues to grow, the integration of robots into everyday tasks requires not only functional, but also social intelligence around human behaviours such as persuasion. This ever-present behaviour is a critical component of human interactions yet has seen minimal research in human-robot interaction (HRI). The objective of this thesis is to understand how people respond to and are influenced by a social robot’s use of a variety of persuasive strategies commonly used in human interactions. This thesis achieves this through a combination of social HRI studies designed to uncover insight about people’s interactions with persuasive robots and through the application of this knowledge into the development of an adaptive learning architecture for robot persuasion. Several studies describe the investigation of a robot’s use of common persuasive dimensions such as rationale for compliance, communication explicitness, sanction use, and authoritative locus of control. These studies highlight the efficacy of emotional persuasive strategies, the importance of less direct persuasive communication, the value of using rewards to encourage compliance, and the challenge of a persuasive robot acting authoritatively.

Building upon these studies, this thesis also describes the development of a hierarchical learning architecture for adaptive persuasion in HRI. This architecture is based upon a well-established psychological framework for how people process persuasive attempts and allows for a robot to adapt its persuasive behaviour to both the static and dynamic factors of a person. The architecture uniquely separates these factors into a hierarchical learning structure, using a contextual bandit approach to optimize a top-level policy for selecting abstract actions that align to static cognitive biases and a Q-learning approach to optimize for the selection of lower-level primitive actions based on a person’s dynamic affective state. Simulated experiments show the improved learning speed and robustness of this architecture over a non-hierarchical benchmark. Between this architecture and the HRI studies performed, this thesis builds an understanding about what makes a robot persuasive and strives to create a more effective adaptive learning approach for use in persuasive HRI, allowing robots to be more effective in real-world, social interactions.