When, What, and How to Explain?

Many mechanisms to explain robot behavior are developed in a vacuum without considering the context of a human–robot interaction. However, when collaborating with a person, additional questions arise concerning explainability: When do explanations become necessary, and what robot modules should they be about?

To answer these questions, I conduct studies establishing when users want explanations from a robot, resulting in a set of scenarios such as robot errors, inabilities, and decision uncertainties. Building on these insights, I develop neurosymbolic agent architectures that leverage foundation models and symbolic planners to provide the right explanation at the right time. Additionally, I explore the use of non-verbal user cues, such as eye gaze, to detect confusion and its source, thereby enabling robots to generate user-centered explanations proactively.

This approach ensures that robot explanations are both relevant and timely, enhancing trust calibration and team performance.