VirtuAI-Safe: Personalized Situated Safety Learning through Integration of VR and LLMs

Abstract Construction safety outcomes are strongly shaped by human factors, including how individuals perceive hazards, interpret contextual cues, and make safety-related decisions in dynamic work environments. Safety education therefore plays a critical role in preparing future construction professionals to reason about risk and apply safety principles in context. Traditional safety instruction often emphasizes regulatory knowledge and procedural compliance, providing limited support for the development of judgment and decision-making skills required on real construction sites. This study examines a personalized situated learning approach for construction safety education that embeds instruction within realistic scenarios while supporting personalized learning. To examine this approach, which integrates virtual reality (VR) and large language models (LLMs), the study implemented VirtuAI-Safe, an immersive virtual learning environment designed to support context-aware and adaptive instructional guidance within simulated construction scenarios. In this study, VirtuAI-Safe is implemented as a construction safety education tool for undergraduate construction management students who are preparing for professional roles in the construction industry. Scaffolding safety was selected as the test domain because it represents a high-risk activity that requires interpretation of evolving site conditions and integration of multiple safety constraints. The system was evaluated using a single-group, postlearning assessment design with undergraduate construction management students. Learning outcomes were assessed using a knowledge test administered 2 days after a 1 h learning session, measuring factual, conceptual, procedural, and transfer knowledge. Learner engagement was measured immediately after the learning session across behavioral, cognitive, and affective dimensions. Overall, the results indicate that VirtuAI-Safe effectively supported learning and engagement within the scope of the study. Postlearning performance was strongest for conceptual understanding and knowledge transfer. Performance on factual and procedural knowledge showed greater variability following the delay, suggesting the need for reinforcement to sustain detailed recall and stepwise execution. Engagement results showed consistently strong behavioral and cognitive engagement, while affective engagement was more moderate. These findings suggest that personalized situated learning environments can support judgment-based aspects of construction safety education and inform future work on adaptive instructional support within immersive learning systems.