Modeling Agent Associations in Human-Robot Interactions
Human-robot interactions are influenced by a dynamic social context that includes humans, robots, and the environment itself. A critical open question in the human-robot interaction community is how we should conceptualize and model the social context to enable desirable robot behavior. In this dissertation, we first contribute a new definition of the term ``social context of a human-robot interaction'' that aims to unite different, prior perspectives on this concept. Critically, we introduce the idea that relationships, or associations, between agents in an interaction are key components of the social context. This dissertation focuses specifically on modeling these associations.
We hypothesize that we can improve computational models of agent associations in a human-robot interaction by leveraging graph abstractions. We incorporate graph abstractions both in data representation and through machine learning models with relational inductive bias. We find that using graph abstractions improves predictive ability for agent associations over methods that do not expressly take advantage of graph structures. Inspired by the natural graph-like structure of social context, we explore this idea from two lenses: human-human and human-robot associations. For each, we use an illustrative predictive task. For human-human interactions we predict conversational groups, and for human-robot interactions we forecast interactions between humans and the robot. To facilitate this research, we utilize data collected from Shutter, the Robot Photographer, a flexible robot platform created to better understand social contexts in dynamic human environments.
Finally, we demonstrate that graph abstractions are useful not just for reasoning about the current or future state of humans in an interaction, but also to shape a robot's behavior. We propose combining action templates with graph neural networks to predict which action a robot should take. This method outperformed other non-graph models while successfully generalizing to previously unseen group sizes.
Overall, this dissertation shows that by leveraging the graph-like structure of the social context, we can create robots that are able to better understand and adapt to the social context in which they operate.