Presentation
Human Factors Evaluation of Gen-AI based NPC Interactions in Public Safety Training
SessionPoster Session 1
DescriptionRecent advances in Generative AI have enabled the development of intelligent, neurobehaviorally diverse Non-Playable Characters (NPCs) to enhance public safety training in extended reality (XR) environments. While LLM-powered NPCs show promise for intelligently driving stressful interactions, critical for de-escalation training, their application in this domain remains largely unexplored. This study evaluates four leading LLMs (Llama3-70b, Claude, Gemini, GPT-4o) for generating AI NPCs with diverse neurobehavioral traits. Sixteen NPCs, based on real-world incidents, were developed using a counterbalanced 2x2 design (cooperative/non-cooperative, neurotypical/ neurodivergent), gender - and race-matched avatars, customized conversational patterns, realistic animations, and AI-generated voices to better simulate the nuances of emergency encounters. Performance was assessed through a web application that tracks decision-making accuracy, behavioral consistency, and user engagement using objective metrics, expert evaluations, and participant feedback. We aimed to identify which LLMs most effectively create believable AI NPCs capable of supporting immersive public safety training. The evaluation will inform the selection of optimal LLMs for designing neurobehaviorally believable AI NPCs that enhance engagement, emotional realism, and training relevance in public safety contexts.
Event Type
Poster
TimeTuesday, October 14th5:30pm - 6:30pm CDT
LocationRiverside East