Abstract
Advancements in the assistive capabilities of robots have led to their increased presence across a variety of domains, ranging from manufacturing to education. As humans collaborate with robots in these environments, it becomes essential to equip robots with the skills necessary to be effective teammates. However, even the most robust robots are bound to experience failures due to hardware limitations, task complexity, or environmental uncertainty. Thus, understanding how such failures affect human-robot interaction is critical for developing resilient systems, and trust provides a means for interpreting their impact. In human-human interactions, trust plays a central role in how teammates delegate tasks, recover from errors, and sustain long-term engagement. Though, trust is a dynamic construct, shaped by both internal and external factors that vary between individuals and contexts. Accurately modeling trust presents significant challenges. Our work investigates how the dynamics in human-human interaction translate to human-robot interaction, and whether these principles can be leveraged to enable robots to recognize real-time changes in human trust and respond appropriately, fostering safer and more resilient collaborations.
First, we examine the implementation of human-like recovery strategies for robots experiencing failures, focusing on sustaining engagement and encouraging future interactions. Building on this, we explore the role of trust in assessing the impact of robot failures, with the understanding that not all failures require the same recovery approach. Next, we create a predictive multimodal model that leverages physiological data to generate real-time predictions of human trust in a robot teammate. Our findings suggest that while trust can be inferred from physiological responses and that robots can employ various strategies to influence human trust, an effective method for calibrating trust is still needed. Specifically, we identify a need to better distinguish factors that constitute trust to enable accurate calibration and to determine how these factors can guide the robot's behavior. To address these gaps, we further break down trust into its cognitive and affective components. In a collaborative escape room scenario, we examine how physiological signals and gaze behaviors relate to these distinct aspects of trust. Using multimodal data, we train a series of models to predict not only overall trust but also cognitive trust (based on competence and reliability) and affective trust (grounded in emotional connection). This more nuanced modeling strategy supports the development of adaptive robots that can respond appropriately to specific trust imbalances and better calibrate human trust during interaction.
