Reproduce “Fusion 360 Gallery: A Dataset and Environment for Programmatic CAD Construction from Human Design Sequences” on OpenResearch

About this paper

Karl D.D. Willis, Yewen Pu, Jieliang Luo, Hang Chu, Tao Du, Joseph G. Lambourne, Armando Solar-Lezama, Wojciech Matusik

Researchers from Autodesk and MIT introduce the Fusion 360 Gallery, the first public dataset of human-designed 3D CAD models explicitly paired with their ground-truth construction sequences, alongside an interactive Fusion 360 Gym environment. This resource enables learning-based approaches to programmatic CAD, achieving up to 67.5% exact reconstruction for test set designs using neurally guided search.

Key insights

• Access to real-world, human-designed programmatic CAD data is essential for training effective learning-based agents, significantly outperforming models trained on synthetic or semi-synthetic data alone.
• Message Passing Networks are well-suited for encoding the rich geometric and topological information within B-Rep face-adjacency graphs, leading to superior performance in predicting CAD operations.
• Programmatic CAD reconstruction, even with simplified action spaces, remains a challenging problem that requires sophisticated learning strategies and search methods to achieve high exact reconstruction rates.

Summary by alphaXiv.