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

A team of researchers has developed a novel framework to solve a critical problem in assistive robotics: the high physical and cognitive burden of personalizing robot behavior for users with profound motor impairments. Traditional methods, like exhaustive pairwise comparisons, are often too fatiguing. The proposed system, detailed in a paper submitted to the 2026 IEEE ROMAN conference, uses Large Language Models (LLMs) as a core translator. It takes unstructured natural language feedback from a user—such as "that move was too fast" or "bring the cup closer"—and converts it directly into executable robotic control code.

To ensure safety and clinical relevance, the LLM is not used in isolation. The pipeline grounds the model's reasoning in the Occupational Therapy Practice Framework (OTPF), a standard clinical model. This step decodes subjective user reactions into explicit physical and psychological needs. These needs are then mapped into transparent, verifiable decision trees that form the robot's policy. Before deployment, an automated "LLM-as-a-Judge" system verifies the structural safety of the generated code.

The framework was validated in a simulated meal preparation study involving 10 adults with paralysis. Results demonstrated that this natural language-based approach significantly reduced user workload compared to traditional preference learning baselines. Furthermore, independent clinical experts reviewed the robot policies generated by the system and confirmed they were both safe and accurately reflected the stated user preferences. This work represents a major step toward making assistive robots more adaptable and user-friendly through intuitive, low-burden communication.