Trust-Aware Multi-Agent Human-Robot Teaming

Abstract

Human-robot teaming is a major emphasis in the ongoing transformation of future workspace wherein human agents and robotic agents are expected to work as a team. Human trust in a robot, a crucial factor for realizing effective human-robot interaction (HRI), has been studied extensively from different angles, including trust's psychological construct, antecedents of trust, and trust calibration. However, the advancements in artificial intelligence and robotics are transitioning robots from subordinates to human partners, which poses new challenges in managing human trust in HRI:

- Previous methods usually assess human trust only at a single point, failing to account for how trust evolves as humans continuously interact with robots. Therefore, there is a critical need to develop models that can dynamically monitor and adapt to changes in trust throughout these interactions. - As robots gain advanced capabilities and are deployed in increasingly complex tasks, determining an optimal trust level that aligns with these scenarios becomes challenging, complicating direct trust calibration efforts. - Previous literature mainly focuses on dyadic HRI scenarios, there is little to no research on trust formation and dynamics in a multi-agent human-robot team. Enabling trust-aware HRI in such scenarios requires a deeper understanding of trust management in multi-agent systems.

To fill the research gap, this dissertation models trust-aware HRI in a computational framework and tackles several fundamental research problems. From Chapter 2 to Chapter 4, we present our results in dyadic human-robot teams. In Chapter 2, we develop a personalized trust prediction model using Bayesian inference, which is built upon previous studies on trust dynamics and outperforms the existing methods in trust prediction. In Chapter 3, we introduce a novel trust-behavior model, the reverse psychology model, and examine the impact of different trust-behavior models on robot policies and team performance. A trust-seeking reward function is proposed to mitigate ``manipulative'' behavior. Chapter 4 looks further into the reward design problem, using the reward shaping technique to balance human trust and task performance, which encourages the robot to increase human trust without significantly compromising task performance.

In Chapters 5 and 6, we extend the problem to multi-agent cases. In Chapter 5, we propose the Trust Inference and Propagation (TIP) model to quantify and predict human trust in multi-human multi-robot teams. This model captures both the direct and indirect experiences a human has with a robot and demonstrates outstanding prediction accuracy through human-subject experiments. Chapter 6 addresses the problem of online multi-agent teaming in a bandit framework. We develop an online learning algorithm, LinMatch, for optimizing robot-human team configurations, achieving efficient matching under uncertainty, and providing novel theoretical bounds. The algorithm's applicability extends beyond HRI to general online matching problems.

By addressing these challenges, this dissertation advances both the theoretical framework and practical applications of trust in human-robot interaction, setting the stage for the development of more reliable and effective autonomous systems in real-world settings.