Low-Burden LLM-Based Preference Learning: Personalizing Assistive Robots from Natural Language Feedback for Users with Paralysis
Researchers have developed a new framework that uses Large Language Models (LLMs) to personalize assistive robots for individuals with paralysis. This system translates unstructured natural language feedback directly into robotic control policies, bypassing the need for traditional, physically demanding preference learning methods. The LLMs are grounded in the Occupational Therapy Practice Framework to interpret user feedback into specific needs, which are then converted into decision trees. An automated LLM-based judge verifies the safety of the generated code before deployment. A study with 10 adults with paralysis demonstrated that this natural language approach significantly reduced user workload compared to existing methods, with occupational therapists confirming the policies' safety and accuracy. AI
IMPACT Enables more intuitive and less burdensome personalization of assistive robots for users with motor impairments.