M:3L Lab

ICLR 2026 Oral

cadrille: Multi-modal CAD Reconstruction with Online Reinforcement Learning

A vision-language model that processes point clouds, multi-view images, and text to generate editable CAD programs. It is the first work to apply online reinforcement learning to CAD reconstruction and achieves a near-zero invalidity ratio.

Publication record

Status
Accepted as an Oral at ICLR 2026
Authors
Maksim Kolodiazhnyi, Denis Tarasov, Dmitrii Zhemchuzhnikov, Alexander Nikulin, Ilya Zisman, Anna Vorontsova, Anton Konushin, Vladislav Kurenkov, Danila Rukhovich
Venue
ICLR 2026 Oral
Research area
CAD Intelligence
First arXiv submission

Abstract / summary

cadrille is a multimodal CAD-reconstruction model that processes point clouds, multi-view images, or text and generates executable Python-based CadQuery programs. It combines supervised fine-tuning with online reinforcement learning driven by programmatic CAD feedback.

Methodology

  • One Qwen2-VL-based vision-language architecture for point clouds, images, and text
  • Supervised fine-tuning on DeepCAD, CAD-Recode, and text-enriched CAD data
  • Executable CadQuery program generation
  • Online Dr. CPPO reinforcement learning with programmatically computed geometric feedback
  • Evaluation by Chamfer Distance, volumetric Intersection over Union, and Invalidity Ratio

Datasets and benchmarks

  • DeepCAD
  • Fusion360
  • CC3D
  • Omni-CAD
  • CAD-Recode
  • Text2CAD

Principal results

  • The paper reports state-of-the-art multimodal CAD-reconstruction results on DeepCAD, Fusion360, and real-world CC3D at publication time.
  • Online RL reduces invalid CAD programs to 0.0-0.2% across the main point-cloud and multi-view-image benchmark rows.
  • On multi-view images, the RL model reports IoU of 92.2% on DeepCAD, 84.6% on Fusion360, and 65.0% on CC3D.
  • On point clouds, the RL model reports IoU of 90.2% on DeepCAD, 85.0% on Fusion360, and 67.9% on CC3D.

Benchmark evidence

BenchmarkTaskInputMetricResultEvidence
DeepCADCAD reconstructionMulti-view images; online Dr. CPPO RLMedian Chamfer Distance x10^3 / IoU % / Invalidity Ratio %0.17 / 92.2 / 0.0cadrille paper, Table 2
Fusion360CAD reconstructionMulti-view images; online Dr. CPPO RLMedian Chamfer Distance x10^3 / IoU % / Invalidity Ratio %0.17 / 84.6 / 0.0cadrille paper, Table 2
CC3DCAD reconstruction from real-world scanned modelsMulti-view images; online Dr. CPPO RLMedian Chamfer Distance x10^3 / IoU % / Invalidity Ratio %0.57 / 65.0 / 0.1cadrille paper, Table 2
DeepCADCAD reconstructionPoint cloud; online Dr. CPPO RLMedian Chamfer Distance x10^3 / IoU % / Invalidity Ratio %0.17 / 90.2 / 0.0cadrille paper, Table 3
Fusion360CAD reconstructionPoint cloud; online Dr. CPPO RLMedian Chamfer Distance x10^3 / IoU % / Invalidity Ratio %0.17 / 85.0 / 0.2cadrille paper, Table 3
CC3DCAD reconstruction from real-world scanned modelsPoint cloud; online Dr. CPPO RLMedian Chamfer Distance x10^3 / IoU % / Invalidity Ratio %0.47 / 67.9 / 0.2cadrille paper, Table 3
DeepCADText-to-CAD reconstructionText; supervised fine-tuningMedian Chamfer Distance x10^3 / IoU % / Invalidity Ratio %0.20 / 82.1 / 1.4cadrille paper, Table 1

Values are attributed to the associated paper and the stated input and evaluation protocol.

Technologies and methods

  • PyTorch
  • PyTorch3D
  • Transformers
  • Qwen2-VL
  • CadQuery
  • Online reinforcement learning

Related M:3L research

Evidence boundary

Benchmark claims and numerical values are paper-reported results for the named protocol. Potential applications are not evidence of completed deployments.