| dc.description.abstract | Robots have the potential to enhance human well-being by assisting with daily activities, particularly for older adults and people with disabilities. One example is robot-assisted dressing, where a robot helps a person put on clothing. However, no two individuals are alike. Each person has unique preferences, behaviors, and needs, making personalization essential for effective assistance. A central challenge is that robots often operate under uncertainty about the human they are helping. This uncertainty may involve the person’s preferences, hidden physical states, or reactions to assistance. If not properly addressed, such uncertainty can lead to ineffective, undesired, or even unsafe outcomes. This thesis asks: How should a robot behave when it is uncertain about the human? To answer this, I present a unified framework for uncertainty-aware personalization in humanrobot interaction, spanning three core components of robot intelligence: preference learning, state estimation, and motion planning. I propose methods that (1) reduce uncertainty using implicit cognitive signals, (2) represent and respect uncertainty through set-based state estimation, and (3) act under uncertainty using relaxed safety constraints. First, I introduce an approach that uses response time, a subtle yet informative cognitive signal, as implicit feedback for preference learning. While traditional methods rely solely on binary choices, I developed the first algorithm that integrates both choices and response times to infer not just what a person prefers, but how strongly they feel about those preferences. Theoretical analysis reveals that response times significantly reduce uncertainty about user preferences, especially when users have strong preferences. In simulation studies, this method decreased misidentification of the most preferred option by up to 55 %, enabling faster and more accurate personalization without extra user input. Second, I address the problem of estimating hidden human states during physical interaction. For example, in dressing, parts of the body, such as the elbow, may be occluded. I introduce the first set-based estimator that represents and respects uncertainty from human behavior and sensing models trained on limited data. Instead of outputting a point estimate, the method constructs a geometric set, such as a 3D box, guaranteed with high probability to contain the true human state. In dressing experiments, the estimator achieved 92 % inclusion using significantly smaller boxes than prior methods, balancing reliability and precision, supporting safe and responsive physical assistance. Third, I consider how a robot should plan motion when it is uncertain about future human behavior. Traditional safety constraints typically prohibit any contact, which can cause the robot to freeze when uncertainty is high. I propose a more flexible definition of safety that allows either collision avoidance or low-impact contact. Integrated into a learning-based control framework, this approach enables efficient motion while maintaining safety. In dressing tasks, it reduced task time by 78 % without compromising safety. Together, these contributions show how robots can reduce, represent and respect, and act under uncertainty to personalize their assistance. This thesis lays a foundation for robots that not only respond to commands, but also understand and adapt to the nuanced, evolving, and uncertain nature of human behavior. | |
