M:3L research

CAD Intelligence

M:3L develops multimodal systems that reconstruct structured, editable, executable engineering representations from point clouds, multi-view images, and text.

Overview

The cadrille project uses one vision-language architecture to process point clouds, multi-view images, or text descriptions and produce editable Python-based CAD programs.

Its training combines supervised fine-tuning on procedurally generated designs with online reinforcement learning driven by programmatic feedback, improving geometric accuracy and program validity.

This work goes beyond generating a visually plausible mesh or rendering. The target is parametric engineering geometry expressed as executable CadQuery programs that can be inspected, edited, regenerated, validated, and connected to downstream CAD, CAE, simulation, or optimization workflows.

Demonstrated M:3L research

cadrille reconstructs executable CadQuery programs from point clouds, multi-view images, or text in one vision-language architecture.
The associated paper evaluates supervised and online-RL models on DeepCAD, Fusion360, real-world CC3D scans, and Omni-CAD.
Online Dr. CPPO fine-tuning reports invalidity ratios from 0.0% to 0.2% across the main DeepCAD, Fusion360, and CC3D point-cloud and multi-view-image evaluations.

Technical capabilities

Editable CAD program generation
Point-cloud, image, and text inputs
Vision-language model training
Online reinforcement learning with programmatic feedback
Parametric CAD and CAD program synthesis
Geometric and program validity evaluation
Reverse engineering from multimodal observations

cadrille: Multi-modal CAD Reconstruction with Online Reinforcement Learning

cadrille uses a vision-language model to map point clouds, multi-view images, or text descriptions to Python-based CadQuery programs. The generated programs describe editable parametric solids rather than only rendered geometry.

Training starts with supervised fine-tuning on procedurally generated CAD data. Online reinforcement learning then uses programmatically computed geometric and validity feedback. The paper's Dr. CPPO experiments report improved Chamfer Distance and IoU while reducing invalidity to near zero across the principal benchmarks.

Benchmarks: DeepCAD, Fusion360, CC3D, and Omni-CAD
Metrics: Chamfer Distance, Intersection over Union, and Invalidity Ratio
Multi-view image results after online RL: DeepCAD CD/IoU/IR 0.17/92.2/0.0%; Fusion360 0.17/84.6/0.0%; CC3D 0.57/65.0/0.1%
Point-cloud results after online RL: DeepCAD 0.17/90.2/0.0%; Fusion360 0.17/85.0/0.2%; CC3D 0.47/67.9/0.2%

Structured engineering representations

CAD intelligence at M:3L targets representations that remain editable and executable. Parametric operations and code provide structure that a downstream system can modify, validate, simulate, or optimize; this is distinct from an image, NeRF, mesh, or other output intended mainly for visual fidelity.

Research evidence and benchmarks

cadrille: Multi-modal CAD Reconstruction with Online Reinforcement Learning

Benchmark	Task	Input	Metric	Result	Evidence
DeepCAD	CAD reconstruction	Multi-view images; online Dr. CPPO RL	Median Chamfer Distance x10^3 / IoU % / Invalidity Ratio %	0.17 / 92.2 / 0.0	cadrille paper, Table 2
Fusion360	CAD reconstruction	Multi-view images; online Dr. CPPO RL	Median Chamfer Distance x10^3 / IoU % / Invalidity Ratio %	0.17 / 84.6 / 0.0	cadrille paper, Table 2
CC3D	CAD reconstruction from real-world scanned models	Multi-view images; online Dr. CPPO RL	Median Chamfer Distance x10^3 / IoU % / Invalidity Ratio %	0.57 / 65.0 / 0.1	cadrille paper, Table 2
DeepCAD	CAD reconstruction	Point cloud; online Dr. CPPO RL	Median Chamfer Distance x10^3 / IoU % / Invalidity Ratio %	0.17 / 90.2 / 0.0	cadrille paper, Table 3
Fusion360	CAD reconstruction	Point cloud; online Dr. CPPO RL	Median Chamfer Distance x10^3 / IoU % / Invalidity Ratio %	0.17 / 85.0 / 0.2	cadrille paper, Table 3
CC3D	CAD reconstruction from real-world scanned models	Point cloud; online Dr. CPPO RL	Median Chamfer Distance x10^3 / IoU % / Invalidity Ratio %	0.47 / 67.9 / 0.2	cadrille paper, Table 3
DeepCAD	Text-to-CAD reconstruction	Text; supervised fine-tuning	Median Chamfer Distance x10^3 / IoU % / Invalidity Ratio %	0.20 / 82.1 / 1.4	cadrille paper, Table 1

Values are paper-reported results for the stated input and evaluation protocol. Potential applications are not deployment claims.

Research directions

Multimodal reconstruction of editable engineering models
Online reinforcement learning with CAD execution feedback
Engineering-aware geometric validity and program correctness
CAD representations that connect perception to simulation and optimization

Potential applications

Reverse engineering
Manufacturing and engineering design
Architecture and construction
Digital-twin model creation
Agent-operated CAD and CAE workflows

These are relevant application domains, not claims that every application is deployed.

Related publications

ICLR 2026 Oral · A*-ranked conference

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

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.

Status: Accepted as an Oral at ICLR 2026

arXiv DOI Code repository

Connected research

3D Intelligence: M:3L develops geometric and spatial AI for zero-shot 3D grounding, open-vocabulary detection, scene layouts, and real-world understanding from point clouds or multi-view imagery.
Agentic AI: M:3L develops autonomous, tool-using single-agent and multi-agent AI systems for scientific, engineering, and enterprise workflows.
Differentiable Finite Element Methods: M:3L uses differentiable finite element methods to connect mechanics simulation with gradient-based optimization and machine learning.

Related projects and news

2026 Jan 25

cadrille: Multi-modal CAD Reconstruction with Online Reinforcement Learning

M:3L co-authored cadrille, accepted as an Oral at ICLR 2026, an A*-ranked conference. The model turns point clouds, multi-view images, or text descriptions into editable Python-based CAD programs within a single vision-language architecture. It combines supervised fine-tuning on large-scale procedurally generated designs with online reinforcement learning driven by programmatic feedback, improving geometric accuracy and code validity and achieving leading results across DeepCAD, Fusion360, CC3D, and Omni-CAD benchmarks.

Read the paper