Dissertation

Link to latest draft of dissertation: Draft as of April 7, 2026

Dissertation Title: Robot Learning During Collaborations with Non-Expert Robot Users

Abstract: As robots integrate into everyday life, they must not only be able to execute a variety of tasks, but must also align their behavior with the varying preferences of non-expert users with whom they collaborate. Because these preferences are often unknown prior to an interaction, robots must be able to learn from non-expert users during collaborations. However, traditional robot learning frameworks often overlook the fact that these collaborations occur within a dynamic social context – a complex interplay of factors related to the agents, their environments, and associations between them.

This dissertation argues that the social context is not merely a backdrop for interaction, but an active mechanism that can be leveraged and shaped to improve robot learning. We contribute a formal taxonomy for the “social context of a human-robot interaction” to provide a unified language for the field. We then interpret and influence the social context in different ways to improve robot learning.

We demonstrate that robots can actively influence the social context to influence how human collaborators provide explicit feedback through the framing and timing of reminders. Then, we show that we can extract additional informative signals that humans “leak” as part of the social context during interactions. We develop models that incorporate this socially contextual information to more accurately predict human preferences over agent behaviors. Additionally, we facilitate combining these types of feedback by contributing the REACT database, a multimodal collection of human reactions and evaluative feedback captured over time during two different interaction scenarios. Finally, we propose a novel mathematical formulation of human preferences over a collaboration and introduce a model to learn from both explicit and implicit human feedback, combining different signals from the social context of the collaboration.

Ultimately, this work demonstrates that by actively engaging with and exploiting the social context of a human-robot interaction, robots can leverage more feedback from non-expert users without increasing human’s teaching burden during human-robot collaborations.