Segmentation-Based Repair of Unstable Procedurally Generated Game Levels in Angry Birds

Abstract

The video game industry has been expanding rapidly, and this huge market means there is always a need for new content. However, the process of creating a game is very time consuming and can take several years. Procedural Content Generation (PCG) is considered one of the solutions to this problem and can increase replay value, reduce production costs, and minimize effort. While PCG methods can generate vast amounts of content efficiently, ensuring structural stability remains a critical challenge, particularly in physics-based games such as Angry Birds. In these games, levels consist of interconnected structures, where even minor instability can lead to unintended collapses. This thesis presents a segmentation-driven repair framework designed to enhance the stability and structural quality of AI-generated game levels for industry use. The proposed framework operates in two main stages: (1) detection of structural gaps using semantic segmentation, and (2) automated reinforcement through targeted block placement. Three segmentation architectures (U-Net, SegFormer, and YOLOv8m-Seg) were trained and evaluated on a dataset of unstable Angry Birds-style levels. Performance was measured using Precision, Recall, F1-score, and Intersection over Union (IoU) for segmentation accuracy, as well as Block Velocity (BV), Block Damage (BD), and Block Destruction (BDes) metrics for stability improvement. Manual evaluation conducted by the author indicated that YOLOv8m-Seg repairs produced the most stable and structurally consistent results to assess perceived stability and structural coherence. Experimental results show that all three models contributed to substantial stability gains, with YOLOv8m-Seg achieving the highest overall performance. Specifically, YOLOv8m- Seg repairs increased the proportion of stable levels up to 45% according to level stability metrics, while maintaining structural consistency. Manual evaluation conducted by the author indicated that YOLOv8m-Seg repairs produced the most stable and structurally consistent results, followed by SegFormer and U-Net. The contributions of this thesis include the development of a modular repair framework for AI-generated levels, a comparative study of segmentation architectures for structural gap detection, and the demonstration of repair strategies that prioritize stability improvements while preserving the original structural design of levels. The proposed framework can be integrated into industry pipelines as a post-processing, designer-assist, or automated QA tool, paving the way for broader adoption of AI-driven content generation in professional game development.re