Making Robot Behaviors Automatically Transparent

Abstract

Incorporating transparency into a robot behavior often demands substantial effort and expertise, to the detriment of anyone who must interact with them closely or for extended periods. We envision a future where this burden is automated away, enabling robot behaviors to be transparent by default. To this end, we propose a conceptual framework that formalizes the distinction between the design and implementation of a behavior and its transparent execution. This separation clarifies the scope of transparency interventions in human-robot interaction (HRI) and highlights key challenges in achieving generalizable transparency. This dissertation addresses these challenges across diverse HRI contexts, demonstrating data- and model-driven augmentations that improve transparency. First, we investigate how users perceive learning robots, uncovering gaps between behavior and external attributions that inform the design of transparency mechanisms. Next, we introduce a data-driven method for controlling the expression of robot affect, enabling behavior creators to balance transparency with task performance. We then present a model-based approach to improving transparency in assistive teleoperation, allowing for explicit control over the trade-off between assistance and intelligibility. Finally, we tackle transparency for human supervisors reviewing robot failures, developing techniques to distill voluminous multi-modal recordings into interpretable summaries. Together, these contributions demonstrate the feasibility of providing transparency into a robot behavior without requiring extensive modification of the original behavior specification. By framing transparency as an independent layer over existing behaviors, this work moves toward automation-friendly, scalable solutions that enhance human-robot interaction across a range of domains.