Date of Award

Spring 1-1-2025

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Computer Science

First Advisor

Vázquez, Marynel

Abstract

Motivated by the value alignment problem, which is concerned with ensuring that an autonomous system's behavior matches user values, this dissertation proposes that creating socially competent mobile robots requires rethinking how success is measured in order to align evaluation metrics with human values and to this end, proposes the use of context-aware simulation systems and subjective human feedback. Social robot navigation is concerned with an agent that must traverse the navigable space in an environment that is shared with people. Conducting such motion in dynamic and densely populated environments demands that a robot understands how humans perceive its behavior and then respond appropriately. This is a challenging task as small deviations in social behavior can significantly impact how people perceive and respond to the robot. Yet, research in social navigation often relies on objective metrics that fail to capture these subtle social factors, leading to policies that may optimize for obvious and easily measurable metrics like efficiency, but neglect the social aspect of how people perceive robot behaviors in terms such as competence. The foundation for this dissertation's contributions is in the area of simulation. We introduce SEAN: the Social Environment for Autonomous Navigation and its follow-on project, SEAN 2.0, a high-visual-fidelity, extensible, and human-centric simulation tool. SEAN allows researchers to develop, test, and compare social navigation algorithms in safe, controlled environments. The contributions that my work makes in the area of simulation are useful for researchers throughout the development lifecycle of social navigation systems. Building on work in simulation, this dissertation makes contributions to evaluating social navigation systems. Fair comparison of existing and future systems allows measurement of and future progress in the field. Critical to fair comparison is a characterization of different social contexts during navigation because social actions are context dependent. Inspired by social psychology, we propose a preliminary set of "Social Situations" that characterize some contexts during social navigation. We then conducted structured interviews with experts working to understand if there is an overarching objective metric which can be used for fair evaluation. We found that beyond safety, the ranking of different metrics varied by application domain. As part of the interviews, we also asked open-ended questions. Responses to these questions highlighted the need to incorporate subjective evaluation criteria, because objective measures alone are insufficient to capture the nuances of the social aspects of navigation. Finally, with the understanding of how critical human perceptions are to the development of social navigation policies, we study the impact of methodological choices researchers can make when collecting human feedback. To enable this work, I led development of the SEAN Experiment Platform (SEAN-EP), which allows researchers to collect human-feedback using interactive, online surveys. Using SEAN-EP, we compare the gold-standard of an interactive, in-person study with scalable online, interactive surveys, and a typical video-based survey. We find that interactive methodologies are preferable to passive alternatives. Still, even with scalable, interactive data collection via SEAN-EP, querying humans for their perceptions of robot behavior requires a significant amount of time and effort. Therefore, we investigate whether it is possible to predict perceptions of robot performance using machine learning in data-limited regimes. Collectively, the contributions of this dissertation provide a foundation for building and evaluating social navigation robots. By integrating context-aware simulation, human-centered evaluation methodologies, and predictive models of subjective human feedback, this work enables more systematic alignment of robot behaviors with people's social expectations. These contributions open the avenue for future research identified in this dissertation, including the development of universally accepted summary metrics for social navigation success, the creation of simulation systems that capture a richer range of human behaviors, and the incorporation of human feedback into robot policies that learn and adapt to predicted human perceptions.

Download

Share

COinS