This study investigates how robots can mediate conflicting human preferences during multi-user navigation. It proposes a three-stage, human-in-the-loop framework that links natural-language negotiation with preference-conditioned reinforcement learning to produce socially compliant motion. The LLM– Human Negotiation Mediator (LLM–HNM) diagnoses conflicts, applies fairness-based decision rules, and translates negotiated compromises into robot trajectories. In a within-subjects study (N=18; six triads), LLM–HNM was compared with two mathematical baselines—a Pareto-optimal mediator and a simple average mediator—under identical conditions. Results show that LLM–HNM achieved safer and more predictable navigation, reducing proxemic intrusions while maintaining high task rewards with only modest efficiency costs. Participants rated it as more dependable, fair, and socially acceptable. The findings demonstrate that fairness-guided, language-based mediation enhances both technical performance and social legitimacy. It contributes to human–robot interaction design by showing that fairness-aware mediation improves group cooperation and acceptance.
Tingting Li (Wed,) studied this question.