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

GameTULearn, a no-code 3D educational game authoring tool, is extended in this work with an AI-assisted scene-synthesis workflow that converts natural-language prompts into editable room layouts. The central thesis is that a language model should be limited to abstract, grid-level semantic decisions, while the game engine handles coordinate placement, validation, and editability within a closed catalogue of educational assets. This division distinguishes the approach from general language-driven scene synthesis systems, which typically draw on open asset spaces, target photorealism or research use, or rely on the model itself for precise spatial reasoning.

In the workflow, a language model selects editor functions and Unity performs execution and validation, so generated scenes remain fully modifiable by educators. A within-subject study with 13 participants found significantly higher perceived usability for the AI-assisted condition than for traditional drag-and-drop authoring (SUS 80.6 versus 69.4, p < 0.001). Qualitative feedback indicated that users valued control and the ability to correct outputs more than full automation, suggesting that assistive rather than autonomous generation is preferred in authoring contexts.

The authors are explicit about the study's constraints. The sample is small, subgroup analyses were not adjusted for multiple comparisons, and the evaluation measures perceived usability rather than educational outcomes. The work is therefore best understood as an early-stage feasibility and usability investigation within a single tool, not as evidence of improved learning. Its motivation, grounded in prior research on game-based learning and barriers to classroom game use, is to lower the technical threshold for educators creating 3D educational games.