Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CADReasoner: Iterative Program Editing for CAD Reverse Engineering

Soslan Kabisov, Vsevolod Kirichuk, Andrey Volkov, Gennadii Savrasov, Marina Barannikov, Anton Konushin, Andrey Kuznetsov, Dmitrii Zhemchuzhnikov

Abstract

Computer-Aided Design (CAD) powers modern engineering, yet producing high-quality parts still demands substantial expert effort. Many AI systems tackle CAD reverse engineering, but most are single-pass and miss fine geometric details. In contrast, human engineers compare the input shape with the reconstruction and iteratively modify the design based on remaining discrepancies. Agent-based methods mimic this loop with frozen VLMs, but weak 3D grounding of current foundation models limits reliability and efficiency. We introduce CADReasoner, a model trained to iteratively refine its prediction using geometric discrepancy between the input and the predicted shape. The model outputs a runnable CadQuery Python program whose rendered mesh is fed back at the next step. CADReasoner fuses multi-view renders and point clouds as complementary modalities. To bridge the realism gap, we propose a scan-simulation protocol applied during both training and evaluation. Across DeepCAD, Fusion 360, and MCB benchmarks, CADReasoner attains state-of-the-art results on clean and scan-sim tracks.

CADReasoner: Iterative Program Editing for CAD Reverse Engineering

Abstract

Computer-Aided Design (CAD) powers modern engineering, yet producing high-quality parts still demands substantial expert effort. Many AI systems tackle CAD reverse engineering, but most are single-pass and miss fine geometric details. In contrast, human engineers compare the input shape with the reconstruction and iteratively modify the design based on remaining discrepancies. Agent-based methods mimic this loop with frozen VLMs, but weak 3D grounding of current foundation models limits reliability and efficiency. We introduce CADReasoner, a model trained to iteratively refine its prediction using geometric discrepancy between the input and the predicted shape. The model outputs a runnable CadQuery Python program whose rendered mesh is fed back at the next step. CADReasoner fuses multi-view renders and point clouds as complementary modalities. To bridge the realism gap, we propose a scan-simulation protocol applied during both training and evaluation. Across DeepCAD, Fusion 360, and MCB benchmarks, CADReasoner attains state-of-the-art results on clean and scan-sim tracks.

Paper Structure

This paper contains 30 sections, 5 equations, 4 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. CAD generation.
  4. Multi-view and point-cloud conditioning.
  5. Iterative refinement and code editing.
  6. Scan-based reverse engineering.
  7. Method
  8. Problem Setup
  9. CADReasoner Architecture
  10. Geometry Encodings
  11. Multi-view image.
  12. Point clouds.
  13. Scan-Simulation
  14. Training
  15. Backbone and single-model setup.
  16. ...and 15 more sections

Figures (4)

  • Figure 1: CADReasoner: scan-to-CAD self-editing loop. (a–c) From a (simulated) scanned mesh we extract target evidence: multi-view images and a point cloud. (d) CADReasoner consumes this evidence together with the current program $C^{t-1}$ and autoregressively predicts an updated CadQuery script $C^{t}$. (e–f) The script is executed to render a new mesh, from which modalities are re-extracted, closing the geometry-driven self-correction loop.
  • Figure 2: Qualitative comparisons on clean surfaces (multi-view, greedy). Nine examples (three each from DeepCAD, Fusion360, MCB). Each panel shows Target, cadrille-SFT, and CADReasoner at $t{=}5$ (best-so-far). Across datasets, CADReasoner produces more accurate reconstructions, recovering small features and correct topology that cadrille often misses or distorts.
  • Figure 3: Multi-view visualization (MCB). We render eight fixed views—six orthographic ($\pm X,\pm Y,\pm Z$) and two isometric—using parallel projection and flat shading; intensity encodes depth. The views are arranged in a $2\times4$ grid with consistent axis orientation.
  • Figure 4: Qualitative evolution on clean surfaces (sampling). CADReasoner’s iterative self-editing under the sampling regime ($N{=}5$ per step). Columns show $t{=}1-5$ and the target. Predictions progressively repair topology and recover small features.