Vitruvion: A Generative Model of Parametric CAD Sketches (2109.14124v2)

Abstract: Parametric computer-aided design (CAD) tools are the predominant way that engineers specify physical structures, from bicycle pedals to airplanes to printed circuit boards. The key characteristic of parametric CAD is that design intent is encoded not only via geometric primitives, but also by parameterized constraints between the elements. This relational specification can be viewed as the construction of a constraint program, allowing edits to coherently propagate to other parts of the design. Machine learning offers the intriguing possibility of accelerating the design process via generative modeling of these structures, enabling new tools such as autocompletion, constraint inference, and conditional synthesis. In this work, we present such an approach to generative modeling of parametric CAD sketches, which constitute the basic computational building blocks of modern mechanical design. Our model, trained on real-world designs from the SketchGraphs dataset, autoregressively synthesizes sketches as sequences of primitives, with initial coordinates, and constraints that reference back to the sampled primitives. As samples from the model match the constraint graph representation used in standard CAD software, they may be directly imported, solved, and edited according to downstream design tasks. In addition, we condition the model on various contexts, including partial sketches (primers) and images of hand-drawn sketches. Evaluation of the proposed approach demonstrates its ability to synthesize realistic CAD sketches and its potential to aid the mechanical design workflow.

• The paper introduces Vitruvion, a generative model that synthesizes CAD sketches by separately modeling geometric primitives and constraints.
• It employs transformer and encoder-decoder architectures to achieve higher log-likelihood and predictive accuracy compared to previous models.
• The approach enables practical functionalities such as autocomplete and sketch inference from hand-drawn images, enhancing CAD design efficiency.

Vitruvion: A Generative Model for Parametric CAD Sketches

In the paper titled "Vitruvion: A Generative Model of Parametric CAD Sketches," the authors propose a novel approach to synthesizing computer-aided design (CAD) sketches using generative modeling. CAD systems are essential tools in various engineering disciplines, enabling the creation of detailed designs through geometric primitives and constraints. The authors introduce Vitruvion as a framework that utilizes machine learning to generate parametric CAD sketches, offering innovations such as autocomplete, constraint inference, and conditional synthesis.

Vitruvion is designed to address the complexities and repetitive nature of CAD design, leveraging the SketchGraphs dataset's real-world examples for training. The model operates by autoregressively synthesizing sketches in terms of primitive sequences followed by constraints that maintain geometric relationships. This approach ensures that sketches generated by Vitruvion can be seamlessly integrated into standard CAD workflows, allowing constraints to propagate edits coherently across sketches.

The authors' methodology includes separate handling of primitives and constraints, with distinct models for each. This division is beneficial as it simplifies implementation by allowing independent training of models while maintaining effective generation of sketches. The primitive model utilizes a transformer architecture to infer sequences of geometric primitives, accommodating scaling and quantization to normalize varying design scales within the dataset. The constraint model, on the other hand, employs an encoder-decoder framework, conditioning on primitives and generating constraints via pointer networks that dynamically adapt to differing numbers of primitives.

Several baselines were tested, demonstrating Vitruvion's capability to synthesize realistic sketches while outperforming models such as those proposed in SketchGraphs and CurveGen. The evaluation of Vitruvion focuses on log-likelihood comparisons, predictive accuracy, and distributional statistics to ensure that generated sketches match real-world designs in terms of complexity and coherence.

The paper also explores conditional generation tasks. Priming the model with incomplete sketches allows it to suggest plausible completions in an autocomplete fashion. Moreover, Vitruvion is extended to infer sketches from images of hand-drawn designs, emulating the early visualization often conducted by engineers. This capability is enabled through sophisticated noise models that adapt the rendering of sketches and employ image encoders resembling vision transformers.

Vitruvion's approach uniquely combines the strengths of machine learning with practical CAD design needs, providing tools that can significantly enhance efficiency in engineering design. The implications of this research extend beyond its immediate applications, hinting at future developments that could include 3D modeling and noise-tolerant syntheses from other types of input data such as 3D scans.

In conclusion, Vitruvion represents a significant advancement in generative modeling for CAD sketches by increasing the fluidity and interactivity of the design process. The results suggest potential improvements in design productivity and quality, fostering broader adoption of AI in mechanical and architectural design fields.