Learning Task Preferences from Real-World Data

Abstract

In order to provide personalized assistance that is capable of adapting to the needs of unique individuals, it is necessary to understand peoples’ preferences for different tasks. Robot assistance often assumes a static model of the individual, while in the real world, people have different capabilities and needs that may change over time. Learning an individual’s task preferences enables the agent to detect when the individual has deviated from their usual behavior, and subsequently understand how to proactively provide assistance when needed. Our work proposes an approach to learn peoples’ preferences for commonplace real-world meal preparation tasks from few demonstrations. We provide two learning methods – mixture-of-experts and meta learning – that condition a model on an individual’s preferences and determine the next step towards completing the task sequence. We evaluate our methods in an in-person user study and data collection with a diverse population of users and real-world kitchen environment on two different tasks. The results highlight the importance of incorporating a representation of users' implicit preferences into personalized predictive models of their behavior.