AI-Assisted Scene Synthesis in the 3D Educational Game Authoring Tool GameTULearn

Abstract

Educational games can improve motivation and learning outcomes but authoring them is still difficult for educators without programming experience. This problem is particularly visible in 3D scene creation, where users must select assets, arrange objects, and maintain spatial consistency inside a complex editor. GameTULearn addresses this challenge as a no-code authoring environment for interactive 3D educational games. This paper presents an AI-assisted scene-synthesis workflow that translates prompt-based natural-language descriptions into editable 3D room layouts inside GameTULearn. The workflow is designed as an assistive layer rather than an automatic generator. A multi-stage pipeline interprets the user request, selects scene-editing functions, retrieves suitable assets and objects, places them on a structured room grid, and executes the resulting operations in Unity. Low-level implementation details such as object instantiation, placement constraints, and validation remain inside the engine. This separation reduces the decision burden for the language model and keeps the generated scene fully editable. We evaluated the workflow in a within-subject user study with 13 participants. Each participant recreated the same educational game scene once with the AI-assisted workflow and once with a traditional drag-and-drop editor. The AI-assisted workflow reached a System Usability Scale score of 80.6, compared with 69.4 for the traditional condition. Participants particularly valued the lower entry barrier and the reduced interaction effort. At the same time, they criticised mismatches between their descriptions and the generated layouts and did not want to give up manual control. The results therefore support a hybrid authoring model in which AI accelerates scene creation, but users remain responsible for review and refinement.