Learning User Preferences using Score-Based Diffusion Mode

Jingze Zhang, Xinyi Yang, Yida Niu, Zhiyuan Zhang

Keywords: diffusion model, object arrangement, utility

Abstract:

Learning user preferences is an essential capability for robots assisting people in
personalized daily tasks. The utility function is commonly used for modeling user
preferences, but learning the utility function is challenging due to the complexity
of human preferences, the context-dependent nature of the utility function, and
the incomplete or inconsistent information from the demonstration. In this work,
we propose a learning framework based on a score-based diffusion model that can
learn utility functions from user demonstrations and can generate instances based
on learned utility functions. We use object arrangement tasks to evaluate utility
learning ability and generalization ability of our model, and achieved significant
improvement compared to existing methods on these tasks. These results show that
the score-based diffusion model is capable of learning user preferences modeled
by utility functions, and can adapt to unseen environment changes.

Submission Number: 234