Priors of U: Dynamic Preference Learning for Embodied Interaction
Abstract
People have their own habits; from how they organize their workspace, over which mug they reach for in the morning, to how they prepare for cooking. Such preferences are context-dependent, dynamic, and often difficult to articulate. As embodied AI agents, such as robots, enter personal spaces, effective interaction requires understanding these implicit and evolving preferences; otherwise, such systems risk being more hindering than helpful. We propose an architecture that learns user preferences from real-world behavioral observation without requiring explicit instruction or feedback. We introduce a factor-graph-based user model that captures preferences across spatial arrangement, tool selection, and workflow, and continuously adapts as tasks and contexts change. The model expands with new observations, enabling open-ended, in situ learning of user-specific preference patterns. Our evaluation shows accurate, calibrated predictions across domains and robust adaptation over time. We demonstrate personalized robot assistance scenarios illustrating seamless and context-aware human–agent interaction. Paper
View paper PDF View in ACM Digital Library
Video Figure