cadrille: Multi-modal CAD Reconstruction with Reinforcement Learning

TL;DR

A multi-modal CAD reconstruction model that simultaneously processes all three input modalities simultaneously is proposed, and the first to explore RL fine-tuning of LLMs for CAD tasks demonstrating that online RL algorithms such as Group Relative Preference Optimization (GRPO) outperform offline alternatives.

Abstract

Computer-Aided Design (CAD) plays a central role in engineering and manufacturing, making it possible to create precise and editable 3D models. Using a variety of sensor or user-provided data as inputs for CAD reconstruction can democratize access to design applications. However, existing methods typically focus on a single input modality, such as point clouds, images, or text, which limits their generalizability and robustness. Leveraging recent advances in vision-language models (VLM), we propose a multi-modal CAD reconstruction model that simultaneously processes all three input modalities. Inspired by large language model (LLM) training paradigms, we adopt a two-stage pipeline: supervised fine-tuning (SFT) on large-scale procedurally generated data, followed by reinforcement learning (RL) fine-tuning using online feedback, obtained programatically. Furthermore, we are the first to explore RL fine-tuning of LLMs for CAD tasks demonstrating that online RL algorithms such as Group Relative Preference Optimization (GRPO) outperform offline alternatives. In the DeepCAD benchmark, our SFT model outperforms existing single-modal approaches in all three input modalities simultaneously. More importantly, after RL fine-tuning, cadrille sets new state-of-the-art on three challenging datasets, including a real-world one. Code is avaliable at https://github.com/col14m/cadrille .

Figures (11)

• Figure 1: Compared to state-of-the-art CAD-Recode, the only existing method that converts point clouds into Python code, cadrille has two key novelties. First, it goes beyond the standard training scheme and adapts LLM RL fine-tuning for CAD reconstruction (left). Moreover, besides point clouds only accepted by single-modal CAD-Recode, cadrille extends to images and textual descriptions, making it the first multimodal approach delivering state-of-art results (right).
• Figure 2: Overview of multimodal data generation pipeline producing textual descriptions, multi-view images and point clouds.
• Figure 3: Overview of cadrille. It can handle three input modalities within a unified framework. Point clouds are processed with a trainable projection layer, while images and texts are passed to a VLM directly. The output of the model is an executable Python script for CAD generation.
• Figure 4: CAD models reconstructed from point clouds from the DeepCAD, Fusion360, and CC3D datasets.
• Figure 5: CAD models reconstructed from multi-view images on the DeepCAD, Fusion360, and CC3D datasets.