CRAFT: Corrective and Robust Multi-Agent Framework for Text-to-Parametric CAD

arXiv:2611.00076v1 Announce Type: cross Abstract: Text-to-CAD generation has the potential to make mechanical CAD more accessible, but existing approaches face a trade-off between data dependence and output fidelity. Supervised methods rely on large paired text-model datasets, while zero-shot prompting strategies are often brittle and tend to lose the parametric structure needed for downstream editing and customization. We present CRAFT, a corrective and robust multi-agent framework that produces parametric OpenSCAD programs from natural language without task-specific training. At its core is a JSON intermediate representation that carries symbolic parametric expressions throughout the generation process, so the resulting models remain editable and retain their parametric structure. CRAFT combines semantic understanding, parametric planning, compilation, rendering, self-correction, and component-level verification within a layered recovery process, using multi-view visual feedback to detect and repair geometric errors. Across NopSCADlib, ABC, and Slice-100K, CRAFT is competitive with direct zero-shot baselines in perceptual and geometric quality while exposing substantially more editable parametric structure.