CNCGEN: A Dataset and Framework for Machining Process Planning and Toolpath Generation from B-rep Models

Xiaolei Zhou; Boyi Lin; Yuchao Feng; Jianwei Zheng

cs.LGSep 30, 2026

CNCGEN: A Dataset and Framework for Machining Process Planning and Toolpath Generation from B-rep Models

Authors: Xiaolei Zhou, Boyi Lin, Yuchao Feng, Jianwei Zheng

Abstract

Learning to generate machining process plans and toolpaths from B-rep CAD requires coupling discrete operation decisions with continuous tool motion as the workpiece evolves. Correctly predicting an operation sequence does not by itself ensure correct material removal, because each toolpath acts on the stock left by preceding cuts. We formulate this problem around persistent manufacturing objects: object identity determines the target of an operation, while the evolving stock state conditions the generation of its toolpath. Based on this formulation, we propose CNCGEN, a dataset and learning framework for three-axis machining. CNCGEN-Dataset contains approximately 50k geometrically verified synthetic machining flows and 800 held-out real CNC records. Each flow aligns B-rep geometry with object-referenced operations, parameterized toolpaths, intermediate stock states, and verification outcomes, enabling supervision of the correspondence between planning decisions and their geometric effects. CNCGEN generates operations and toolpaths for selected objects step by step, updating a compact machining state to guide subsequent predictions. During training, a learned surrogate verifier provides material-removal feedback that links local predictions to their geometric consequences. Experiments on synthetic and held-out real CNC records show that CNCGEN improves the resulting workpiece geometry and reduces residual material and overcut compared with adapted CNC generation baselines.

Figures & tables

Fig. 1: Overview of CNCGEN . Given a B-rep model, CNCGEN identifies manufacturing objects and generates a machining plan with corresponding toolpaths. The highlighted objects and steps illustrate how a shared identity links discrete operation decisions to continuous tool motion. The upper loop represents feedback of the updated state to subsequent operations and toolpath predictions.

Fig. 2: CNCGEN-Dataset samples. Wireframe B-rep views are paired with rendered machining-feature geometry and verified machining outcomes. Fig. S2 provides additional examples.

Fig. 3: Architecture of CNCGEN . The four panels show feature recognition, operation planning, toolpath generation, and verification-guided training. Machining-state feedback links toolpath generation with prefix refinement. The Surrogate Verifier is frozen during generator refinement.

Method	MRP source	IoU $\uparrow$	Removal F1 $\uparrow$	CD $\downarrow$	Overcut $\downarrow$	Residual $\downarrow$
CNC-Net+Pred. MRP	Predicted	0.7144	0.2318	0.0595	0.2327	0.6032
CNC-Net+Oracle MRP	Oracle	0.9271	0.5067	0.0309	0.0025	0.5973
CNCGEN (Ours)	N/A	0.9525	0.9016	0.0268	0.0079	0.0431

Table 1: Synthetic-test geometry quality after execution. CNC-Net+Pred. MRP uses predicted region priors; Oracle MRP uses annotation-derived priors as a diagnostic setting. CD is normalized by the stock bounding-box diagonal. Bold marks the best value in each column.

Variant	Step EM $\uparrow$	Seq. EM $\uparrow$	Macro-F1 $\uparrow$	IoU $\uparrow$	Removal F1 $\uparrow$	CD $\downarrow$	Overcut $\downarrow$	Residual $\downarrow$
Full CNCGEN (Ours)	0.8552	0.8290	0.9024	0.9525	0.9016	0.0268	0.0079	0.0431
w/o MER	N/A †	N/A †	0.2182	0.9315	0.1484	0.0317	0.0086	0.0673
w/o prefix refiner	0.8172	0.7796	0.8954	0.9332	0.3115	0.0274	0.0167	0.0475
w/o state-aware toolpaths	0.8530	0.8260	0.9012	0.9327	0.3113	0.0273	0.0169	0.0477
w/o SV loss	0.8400	0.8132	0.8778	0.9296	0.3041	0.0286	0.0236	0.0472

Table 2: Ablation results on CNCGEN-Dataset . Each variant disables one component of CNCGEN while preserving the same evaluation interface. Bold marks the best value in each column.

Fig. 4: Geometric outcomes on held-out real CNC records. Left: representative B-rep models, feature geometry, and machining outcomes from the benchmark. Right: target geometry and final geometries produced by CNCGEN and CNC-Net+Pred. MRP under the offline evaluator.

Method	MRP source	IoU $\uparrow$	Removal F1 $\uparrow$	CD $\downarrow$	Overcut $\downarrow$	Residual $\downarrow$
CNC-Net+Pred. MRP	Predicted	0.6275	0.1170	0.0703	0.3205	0.7268
CNC-Net+Oracle MRP	Oracle	0.9203	0.4290	0.0332	0.0112	0.6617
CNCGEN (Ours)	N/A	0.9274	0.8126	0.0148	0.0134	0.1845

Table 3: Geometry quality on 800 held-out real CNC records, evaluated with the same near-net-stock protocol as the synthetic test set. Pred. MRP uses predicted region priors, while Oracle MRP uses annotation-derived priors. CD is normalized by the stock bounding-box diagonal. Bold marks the best value in each column.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Fig. S1: CNC machining primitives and execution representations used by CNCGEN-Dataset and CNCGEN : B-rep input, extracted machining feature, recorded toolpath, and operation primitives for slant milling, pocket milling, and drilling.

Fig. S2: Extended gallery of twelve CNCGEN-Dataset examples. Each example pairs a wireframe B-rep model with rendered manufacturing-feature geometry and a verified machining outcome. Fig. 2 shows a subset.

Fig. S3: Representative B-rep parts from the held-out real CNC benchmark, illustrating machining geometries beyond the synthetic standard-part families.

Fig. S4: Synthetic dataset statistics: distribution of feature counts and proportions of command types in the generated machining flows.

Fig. S5: Aligned training targets and verification supervision. Representative machining flows align B-rep geometry, machining-feature targets, toolpath/process targets, and verified execution outcomes; the verifier supervision compares generated stock evolution with the target geometry.

Target	MAE $\downarrow$	RMSE $\downarrow$	Spearman $\rho$ $\uparrow$
IoU	0.048	0.071	0.79
Target preservation	0.021	0.035	0.61
Overcut penalty	0.006	0.019	0.76
Residual penalty	0.056	0.083	0.74

Appendix

Table S1: Held-out SV agreement with continuous geometric targets. MAE and RMSE measure prediction error; Spearman correlation measures ranking agreement.

Target	AUROC $\uparrow$	Precision $\uparrow$	Recall $\uparrow$	F1 $\uparrow$	Acc. $\uparrow$
Risk label	0.958	0.902	0.887	0.894	0.901
Validity label	0.951	0.889	0.932	0.910	0.904

Appendix

Table S2: Held-out SV risk and validity screening.

Fig. S6: Geometric outcomes on held-out real CNC parts. Rows show target geometry, CNC-Net+Pred. MRP, and CNCGEN , from top to bottom, under the common offline evaluator.

Fig. S7: Representative failed rollouts. Target geometries in the top row are compared with generated results in the bottom row, showing residual material, overcut, and errors near feature boundaries.

Explore similar work

BenchCAD: A Comprehensive, Industry-Standard Benchmark for Programmatic CAD
May 11, 2026Haozhe Zhang, Kaichen Liu, Miaomiao Chen +4Parametric Computer-Aided Design Programs Image-To-Code Generation
University of Virginia · University of California, San Diego · Rice University
CADENA: Stepwise CAD Reverse Engineering
Aug 1, 2026Soslan Kabisov, Gennadiy Savrasov, Maksim Elistratov +9Parametric Computer-Aided Design Programs 3D Models
Lomonosov Moscow State University · DAIMLD · Independent Researcher +1
ArtisanCAD: An Industrial-Level CAD Agent with Expert-Grounded Knowledge Distillation
Jul 7, 2026Yunhan Xu, Qifeng Wu, Xunjin Li +10Image-To-Cad Generation Procedural
Peking University, Beijing 100871, China · TenFong Technology Co., Ltd., Shenzhen 518000, Guangdong, China · Eastern Institute of Technology, Ningbo 315200, Zhejiang, China +2

May 11, 2026, cs.AIJ/K move · Enter open · S save

BenchCAD: A Comprehensive, Industry-Standard Benchmark for Programmatic CAD

Haozhe Zhang, Kaichen Liu, Miaomiao Chen, Lei Li+3

University of Virginia · University of California, San Diego · Rice University

Parametric Computer-Aided Design Programs Image-To-Code Generation 3D Models Industrial

Open paper AskPDF arXiv

Industrial Computer-Aided Design (CAD) code generation requires models to produce executable parametric programs from visual or textual inputs. Beyond recognizing the outer shape of a part, this task involves understanding its 3D structure, inferring engineering parameters, and choosing CAD operations that reflect how the part would be designed and manufactured. Despite the promise of Multimodal large language models (MLLMs) for this task, they are rarely evaluated on whether these capabilities jointly hold in realistic industrial CAD settings. We present BenchCAD, a unified benchmark for industrial CAD reasoning. BenchCAD contains 17,900 execution-verified CadQuery programs across 106 industrial part families, including bevel gears, compression springs, twist drills, and other reusable engineering designs. It evaluates models through visual question answering, code question answering, image-to-code generation, and instruction-guided code editing, enabling fine-grained analysis across perception, parametric abstraction, and executable program synthesis. Across 10+ frontier models, BenchCAD shows that current systems often recover coarse outer geometry but fail to produce faithful parametric CAD programs. Common failures include missing fine 3D structure, misinterpreting industrial design parameters, and replacing essential operations such as sweeps, lofts, and twist-extrudes with simpler sketch-and-extrude patterns. Fine-tuning and reinforcement learning improve in-distribution performance, but generalization to unseen part families remains limited. These results position BenchCAD as a benchmark for measuring and improving the industrial readiness of multimodal CAD automation.

Abstract

Explore similar work

BenchCAD: A Comprehensive, Industry-Standard Benchmark for Programmatic CAD

CADENA: Stepwise CAD Reverse Engineering

ArtisanCAD: An Industrial-Level CAD Agent with Expert-Grounded Knowledge Distillation