FRANC: Feeding Robot for Adaptive Needs and Personalized Care

Abstract

Robot-assisted feeding systems have the potential to significantly enhance the independence and quality of life of individuals with mobility impairments. While prior work has focused on personalizing bite sequences based on user feedback provided only at the start of the feeding process, this approach assumes that users can fully articulate their preferences upfront. In reality, it is cognitively challenging for users to anticipate every detail, and their preferences may evolve during feeding. Thus, there is a need for an adaptive system that supports iterative corrections across all stages of the feeding process while maintaining context and feeding history to interpret inputs relative to earlier instructions. In this paper, we present FRANC, a novel framework for personalized RAF that leverages large language models (LLMs) with a decomposed prompting strategy to dynamically adjust bite sequence, acquisition and transfer parameters during feeding. Our approach allows iterative corrections without sacrificing consistency and accuracy. In our user studies, FRANC improved bite sequencing accuracy from 65% to 93% and enhanced user satisfaction, with participants reliably perceiving when their preferences were being integrated despite occasional execution failures. We also provide a detailed failure analysis and offer insights for developing more adaptive and effective robot-assisted feeding systems.