Hierarchical Multi-Level 3D Geometry Generation with Stress-Aware Learning

TMLR Paper6341 Authors

Abstract: Current approaches for Lego 3d structural assembly are usually learned to maximize IOU between generated output and target construction. We propose a new approach which is able to build stable structures based on physics-aware reward. Our method employs a two-level agent architecture in which a high-level PPO-based planner proposes a scheme, while a low-level Wave Function Collapse (WFC) agent handles precise brick placement with constraint satisfaction. Experimental results demonstrate that our hierarchical method consistently constructs structurally sound buildings while reducing material usage. We also show that replacing the computationally expensive FEM solver with fast FNO achieves comparable performance, confirming the approach's scalability for large-scale problems.

Submission Type: Long submission (more than 12 pages of main content)

Assigned Action Editor: ~Weijian_Deng1

Submission Number: 6341