LuBan: Constructing Hierarchical Graphs for CAD Sketch Generation via Transformer Intermediate Outputs

Area of Science:

• Computer Science
• Artificial Intelligence
• Computer-Aided Design

Background:

• Computer-Aided Design (CAD) sketches are crucial for industrial design and manufacturing, built from geometric primitives and constraints.
• Current AI methods often process primitives and constraints separately, missing their inherent connections.
• This limits the efficiency and accuracy of automated CAD sketch generation from various inputs.

Purpose of the Study:

• To introduce LuBan, an end-to-end AI model for generating CAD sketches.
• To overcome limitations of existing methods by jointly learning primitives and constraints.
• To enhance the design process through more intuitive and accurate sketch conversion.

Main Methods:

• Utilized the DEtection TRansformer (DETR) architecture for primitive modeling.
• Differentiated between parametric and non-parametric features within the model.
• Derived sub-primitive features to predict constraints, capturing primitive-constraint relationships.
• Represented generated CAD sketches as hierarchical graphs.

Main Results:

• LuBan achieved state-of-the-art performance on both precise and hand-drawn renderings.
• Demonstrated superior effectiveness compared to models trained on primitives and constraints independently.
• Validated through qualitative and quantitative experiments and ablation studies.

Conclusions:

• LuBan effectively models the intrinsic relationships between primitives and constraints in CAD sketches.
• The end-to-end approach simplifies and accelerates the CAD sketch generation process.
• The model aligns with the "what you draw is what you get" principle, benefiting designers.