VR-AI lab safety training: a virtual reality game with AI-powered guidance for hazardous scenario response

Abstract

This report presents the design and implementation of a single-player Virtual Reality (VR) training system enhanced with Artificial Intelligence (AI) for immersive and adaptive laboratory safety education. Traditional methods of lab safety instruction often lack interactivity and fail to simulate real-world urgency, limiting student engagement and retention. To address these gaps, we developed a Unity-based VR application that simulates hazardous lab scenarios—such as chemical fires—allowing users to interact with virtual equipment and perform safety protocols in real time. An AI assistant powered by GPT-4 is integrated via LangChain, offering context-sensitive, natural language guidance during training. Semantic caching with Redis optimizes response latency and reduces redundant API calls by storing and retrieving AI responses based on game state and query embeddings. A Rust-based backend handles session management and scoring persistence through PostgreSQL, ensuring scalability and clean data separation. The system emphasizes modularity through dedicated Unity managers for transitions, game logic, and API interaction. Scoring is dynamically calculated based on hazard resolution time and completion of weighted safety subtasks. Initial validation was conducted through manual testing of gameplay mechanics, interaction reliability, and AI response behavior. This work contributes a modular, scalable framework for AI-guided VR safety education and demonstrates how experiential learning models, supported by intelligent tutoring systems, can enhance training in safety-critical environments. Future directions include expanding to multiplayer, introducing voice-based AI queries, and increasing the fidelity of lab equipment simulations.