Low-Burden LLM-Based Preference Learning: Personalizing Assistive Robots from Natural Language Feedback for Users with Paralysis

The article presents a novel offline framework for personalizing assistive robots using natural language feedback, aimed at reducing cognitive and physical strain for users with paralysis. This approach employs Large Language Models (LLMs) to translate unstructured feedback into deterministic robotic control policies, which are then verified for safety by an "LLM-as-a-Judge." Validation in a simulated meal preparation study indicates that this method significantly decreases user workload while ensuring that the generated policies align with user preferences, highlighting its potential for enhancing user experience in assistive robotics.