Exploring Passenger-Automated Vehicle Negotiation Utilizing Large Language Models for Natural Interaction

As vehicle automation advances to SAE Levels 3 to 5, transitioning driving control from human to system, ensuring automated vehicles (AVs) align with user preferences becomes a challenge. Natural interaction emerges as a common goal, offering ways to convey user interests in a user-friendly manner. However, technical, legal, or design constraints may prevent fulfilling these preferences, leading to potential conflicts. Through an online survey (N=50), potential driver-passenger conflicts and their handling strategies were explored. Subsequently, in a Virtual Reality study (N=14), we applied identified strategies (ranging from distracting to motivating and adhering to social norms) to user-AV interactions using a state-of-the-art language model (GPT-4 Turbo) primed with the strategies to simulate realistic dialogues. Additionally, adaptive communication was compared to non-adaptive communication. Our findings reveal a preference for adaptive communication. Yet, despite using advanced modeling, accurately predicting user interactions remained challenging, with users often trying to outsmart the AI.