AI-directed procedural content generation for personalized XR in industrial-heritage museums

Abstract

This paper introduces an AI-directed procedural content generation (PCG) framework for Extended Reality (XR), implemented as a rule-based expert system with probabilistic scoring, and describes its use in an industrial heritage museum. The system uses visitor behavior to guide the progression of experience and architectural changes. By analyzing real-time movement, head orientation, and hand movements, an AI Director identifies one of four interaction profiles (Observer, Shaper, Explorer, Disruptor) and adjusts the tasks and feedback accordingly. Adaptation is limited by museum semantics at the space, collection, and artifact levels to maintain curatorially consistent XR experiences. The framework is implemented and evaluated at the santralistanbul Energy Museum, with a four-scene XR experience (tutorial, profiling, and two adaptation scenes) on a standalone headset. A multimodal logging system was implemented to monitor user behavior and support a single-group deployment study with 37 participants. Telemetry-derived profile vectors showed agreement with self-reported profiles in this exploratory sample (tie-aware accuracy: 83.8%, 95% Wilson CI [68.9%, 92.3%]; strict-subset single-label accuracy: 70.8%, 95% Wilson CI [50.8%, 85.1%]). Pre/post questionnaires also showed higher industrial-heritage perception after the XR session than before it (4.61 vs. 5.96 on a one to seven scale; mean paired change = 1.35, t(36) = 11.54 , p < . 001 , d(z) = 1.90). Immersion stays high, workload is moderate, and usability is high. Heritage-change scores are positively correlated with immersion and usability, but not with workload. These patterns are consistent with the feasibility of curatorially constrained, profile-aware PCG in an industrial-heritage XR setting.