Reproduce “ReCAD: Reinforcement Learning Enhanced Parametric CAD Model Generation with Vision-Language Models” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Jiahao Li, Yusheng Luo, Yunzhong Lou, Xiangdong Zhou

The ReCAD framework combines supervised fine-tuning and reinforcement learning, using parameterized CAD code as guidance, to produce precise and editable parametric CAD models from multimodal inputs. This method achieves improved geometric accuracy and generalization across text-to-CAD and image-to-CAD generation tasks, outperforming existing state-of-the-art approaches.

Key insights

  • Integrating RL with PLMs, guided by parameterized code, effectively enhances reasoning and generalization for complex, constraint-driven domains like CAD, moving beyond SFT limitations.
  • Representing CAD models hierarchically using functional interfaces for parameterized code improves flexibility, precision, and editability compared to direct low-level CAD sequence generation.
  • A unified reward function that combines geometric (IOU_best), visual (DINOv2 similarity), and semantic/formatting fidelity is critical for reliably guiding RL towards high-quality CAD generation.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.