IterCAD: An Iterative Multimodal Agent for Visually-Grounded CAD Generation and Editing

Computer-Aided Design (CAD) is essential in modern manufacturing, yet traditional automated methods typically operate in a one-shot manner, which does not reflect the iterative processes used by designers. IterCAD is developed as a multimodal agent framework that facilitates closed-loop interactions for CAD generation and editing, addressing three key tasks: Drawing-to-Code, Text-to-Code, and Interactive Editing.

A data synthesis pipeline was created to generate compliant multi-view engineering drawings and complex code-editing tasks. The agent is optimized using progressive supervised fine-tuning (SFT) and geometry-aware reinforcement learning, which improves both code executability and geometric fidelity.

The introduction of the IterCAD-Bench evaluation suite and the Chamfer Distance Tolerance-Recall (CD-TR) curve provides a new standard for assessing code validity and geometric precision. Results indicate that IterCAD achieves superior performance across multiple benchmarks, demonstrating its effectiveness in iterative refinement processes.

This advancement has significant implications for engineers seeking to enhance their CAD workflows.