Personalization in Human-Robot Interaction through Preference-based Action Representation Learning

• DOI: 10.48550/arxiv.2409.13822
• Publication Date: 2024-09-20
• Authors: Ruiqi Wang; Dezhong Zhao; Dougyoung Suh; Ziqin Yuan; Guohua Chen; Byung‐Cheol Min
• Keywords: Representation; Preference learning; Human–robot interaction

Preference-based reinforcement learning (PbRL) has shown significant promise for personalization in human-robot interaction (HRI) by explicitly integrating human preferences into the robot process. However, existing practices often require training a personalized policy from scratch, resulting inefficient use of feedback. In this paper, we propose preference-based action representation (PbARL), an efficient fine-tuning method that decouples common task structure preference leveraging pre-trained policies. Instead directly with preference, PbARL uses it as reference maximizes mutual information between source domain and target user preference-aligned domain. This approach allows to personalize its behaviors while preserving original performance eliminates need extensive prior domain, thereby enhancing efficiency practicality real-world HRI scenarios. Empirical results on Assistive Gym benchmark study (N=8) demonstrate benefits our compared state-of-the-art approaches.