Research map / 3D vision / 3D generation
CNC-Net: Self-Supervised Learning for CNC Machining Operations
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes CNC-Net, a self-supervised DNN that simulates CNC milling, drilling and rotation operations to carve a bounding-box material into a target 3D CAD shape without operation labels.
Problem
Manual CNC programming is labor-intensive and expert-dependent, and no dataset provides ground-truth sequences of CNC machining operations for supervised learning.
Approach
Model CNC milling/drilling with cylindrical tool primitives and CSG subtraction from a bounding-box stock, with learned tool radius, toolpath, drill tip and workpiece rotation, trained end-to-end with self-supervised voxel losses rather than operation labels.
Key results
- On ABC, CNC-Net reaches IoU 0.824 vs 0.787 for the best baseline (CSG-Stump), but has worse CD (0.509) and NC (0.893) than CAPRI-Net (0.312 CD, 0.914 NC).
- On ShapeNet, CNC-Net achieves IoU 0.740 vs 0.700 for CAPRI-Net, but CD 1.562 is worse than CAPRI-Net's 0.447.
- Zero-shot per-shape training gives IoU 0.780 on ABC, versus 0.824 with pre-training plus fine-tuning.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CNC-Net instead of the strongest reported CAD reconstruction baseline | 3D CAD reconstruction · ABC | IoU: 0.787 -> 0.824 | Helped |
| Use CNC-Net instead of the strongest reported CAD reconstruction baseline | 3D CAD reconstruction · ShapeNet | IoU: 0.700 -> 0.740 | Helped |
| Use CNC-Net instead of the strongest reported CAD reconstruction baseline | 3D CAD reconstruction · ABC | CD: 0.312 -> 0.509 | Hurt |
| Use CNC-Net instead of the strongest reported CAD reconstruction baseline | 3D CAD reconstruction · ShapeNet | CD: 0.447 -> 1.562 | Hurt |
| Train individually in a zero-shot manner instead of fine-tuning (ablation) | CNC-Net · ABC | IoU: 0.824 -> 0.780 | Hurt |
| Train individually in a zero-shot manner instead of fine-tuning (ablation) | CNC-Net · ShapeNet | IoU: 0.740 -> 0.698 | Hurt |
| Add milling operations (ablation) | CNC-Net · ABC | IoU: 0.460 -> 0.780 | Helped |
| Add drilling operations (ablation) | CNC-Net · ABC | IoU: 0.525 -> 0.780 | Helped |
| Add rotation operations (ablation) | CNC-Net · ABC | IoU: 0.776 -> 0.780 | Helped |
| Add the shape loss and center loss alongside the essential losses (ablation) | CNC-Net · ABC | reconstruction performance | Helped |
Surprising findings
- Removing the drilling operation roughly doubles the CD (from 1.127 to 2.098 when rotation is kept and drilling removed in the ablation table), though it only modestly changes IoU.
- The carving approach yields higher volumetric IoU but markedly worse surface metrics (CD and NC) than union-based primitive methods.
Limitations
- No ground-truth operation sequences exist, so optimality of learned operations cannot be assessed.
- Only cylindrical tools and bounding-box stock are modeled; only milling, drilling and rotation are supported.
- Surface-based metrics (CD, NC) are worse than the union-based baselines because carving produces less smooth shapes.
Experimental setup
Datasets: ABC ShapeNet Core (V1)
Benchmarks: ABC ShapeNet
Metrics: IoU F1 Chamfer Distance Normal Consistency
Baselines: CSG-Stump ExtrudeNet CAPRI-Net SECAD-Net
Builds on: CAPRI-Net CSG-Stump ExtrudeNet SECAD-Net
Closest papers by meaning
- CADCrafter: Generating Computer-Aided Design Models from Unconstrained Images (CVPR 2025 · similarity 0.80)
- MiCADangelo: Fine-Grained Reconstruction of Constrained CAD Models from 3D Scans (NeurIPS 2025 · similarity 0.80)
- Neural Assembler: Learning to Generate Fine-Grained Robotic Assembly Instructions from Multi-View Images (AAAI 2025 · similarity 0.79)
- Mamba-CAD: State Space Model for 3D Computer-Aided Design Generative Modeling (AAAI 2025 · similarity 0.79)
- StepCAD: Mesh-to-CAD Code Generation via LLM Policy and Geometry-Guided Search (NeurIPS 2026 · similarity 0.79)
- Self-Supervised Dual Contouring (CVPR 2024 · similarity 0.78)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.