cadrille: Multi-modal CAD Reconstruction with Online Reinforcement Learning

News:

• 🔥 May, 2025. cadrille is state-of-the-art in three CAD reconstruction benchmarks:
  DeepCAD
  Fusion360
  CC3D
  

This repository contains an implementation of cadrille, a multi-modal (point clouds / images / text) 3D CAD reconstruction method introduced in our paper:

cadrille: Multi-modal CAD Reconstruction with Online Reinforcement Learning
Maksim Kolodiazhnyi, Denis Tarasov, Dmitrii Zhemchuzhnikov, Alexander Nikulin, Ilya Zisman, Anna Vorontsova, Anton Konushin, Vladislav Kurenkov, Danila Rukhovich
AIRI Institute
https://arxiv.org/abs/2505.22914

Installation

Install Python packages according to our Dockerfile. We support DeepCAD (test), Fusion360 (test), Text2CAD (train / val / test), and CAD-Recode (train, val) datasets. Follow our instruction to download and preprocess data.

Train

To start training run train.py script:

python train.py --mode pc_img --use-text

To disable some of the modalities set --mode to img or pc, or disable --use-text. We don't provide RL fine-tuning code for now. Alternatively both SFT and RL models can be downloaded from 🤗 HuggningFace.

Inference

To predict CadQuery codes run test.py script:

python test.py --split deepcad_test_mesh --mode pc

To run on other datasets and modalities use --split fusion360_test_mesh or set --mode to img or text.

Evaluation

To evaluate IoU, invalidity ratio, and chamfer distance run evaluate.py script:

python evaluate.py

Citation

If you find this work useful for your research, please cite our paper:

@article{kolodiazhnyi2025cadrille,
  title={cadrille: Multi-modal CAD Reconstruction with Online Reinforcement Learning},
  author={Maksim Kolodiazhnyi, Denis Tarasov, Dmitrii Zhemchuzhnikov, Alexander Nikulin, Ilya Zisman, Anna Vorontsova, Anton Konushin, Vladislav Kurenkov, Danila Rukhovich},
  journal={arXiv preprint arXiv:2505.22914},
  year={2025}
}

Acknowledgments

This work was supported by Artificial Intelligence Research Institute