Learning Hierarchical and Geometry-Aware Graph Representations for Text-to-CAD
Summary: arXiv:2604.10075v1 Announce Type: new
Abstract
Text-to-CAD code generation represents a significant challenge in translating textual instructions into long sequences of interdependent operations. Traditional methods often decode text directly into executable code, such as bpy, without accounting for the complexity of assembly hierarchy or geometric constraints. This oversight results in an expanded search space, leads to the accumulation of local errors, and frequently triggers cascading failures during the assembly of complex structures.
Proposed Solution
To tackle these challenges, we introduce a hierarchical and geometry-aware graph as an intermediate representation for the text-to-CAD task. This graph models multi-level parts and components as nodes while encoding explicit geometric constraints as edges. Our innovative framework does not simply map text to code; instead, it first predicts the necessary structure and constraints, which then informs the sequencing of actions and the generation of code. This methodological shift enhances both geometric fidelity and the satisfaction of geometric constraints.
Curriculum Learning Strategy
A key component of our approach is the implementation of a structure-aware progressive curriculum learning strategy. This strategy constructs graded tasks through controlled structural edits, thereby allowing us to explore the boundaries of the model’s capabilities. Additionally, it synthesizes boundary examples for iterative training, which helps refine the model further and improve its performance.
Dataset and Evaluation Metrics
To support our framework, we have developed a comprehensive dataset consisting of 12,000 entries that feature instructions, decomposition graphs, action sequences, and corresponding bpy code. Alongside this dataset, we introduce graph- and constraint-oriented evaluation metrics that provide a robust framework for assessing model performance in a meaningful way.
Experimental Results
Our extensive experiments demonstrate that the proposed method consistently outperforms existing approaches in two critical areas: geometric fidelity and the accurate satisfaction of geometric constraints. The results highlight the effectiveness of employing hierarchical and geometry-aware representations, showcasing the potential for significant advancements in text-to-CAD code generation.
Conclusion
The introduction of hierarchical and geometry-aware graph representations marks a pivotal step forward in the field of text-to-CAD code generation. By addressing the limitations of traditional methods and enhancing the model’s ability to understand and manipulate geometric constraints, we pave the way for more accurate and reliable CAD generation from textual input. This research not only contributes to the academic field but may also have practical implications in various industries that rely on CAD technologies.
Future Work
- Exploring additional geometric constraints and their implications on assembly processes.
- Integrating machine learning techniques to further enhance model accuracy and efficiency.
- Expanding the dataset to include a wider variety of instructions and CAD scenarios.
- Investigating real-world applications and partnerships to validate the framework’s utility.