acceptodds
Under review as a conference paper at ICLR 2027

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

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 introduce 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 CNC-Net adapted with predicted machining-region priors.

Primary area:
applications to robotics, autonomy, planning
Keywords:
CAD-to-CAM learning, machining process planning, toolpath generation, B-rep, object-centric learning, datasets and benchmarks

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_rYpTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the joint generation of machining process plans and continuous toolpaths from B-rep models. It represents machining as a state-evolution process organized around persistent manufacturing objects: process planning and toolpath generation share object identities, while subsequent operations are predicted based on the stock state left by previous cuts.

Critical strengths

The problem is practically relevant, and the formulation reflects the nature of machining. Operation selection and toolpath generation are closely linked: a reasonable operation sequence does not guarantee correct material removal, because each subsequent toolpath acts on the stock left by earlier cuts. Connecting discrete operation decisions, continuous tool motion, and state updates is therefore well motivated, particularly when machining regions overlap.

The dataset provides comprehensive process-level supervision and has potential value as a research resource. The dataset includes not only target geometry and operation labels, but also aligned manufacturing objects, toolpaths, intermediate stock states, and verification outcomes. These records are useful for studying how machining decisions affect the resulting geometry. The 800 real CNC records are reserved for testing and are not used for training, validation, or fine-tuning, providing a useful setting for assessing transfer from synthetic training data.

Critical weaknesses

The method appears to be an integration of existing techniques, and its core novelty is not yet convincing. As the paper itself discusses, CNC-Net already updates the stock after each operation, and DeepMS uses intermediate states to predict machining sequences. Accounting for material removed by earlier operations is therefore not a new idea introduced in this paper. CNCGEN further connects object representations, operation prediction, and toolpath generation. This is a reasonable design, but I do not see a particularly clear methodological advance.

The baseline comparison is too limited to make the large performance gains fully convincing. The main experiments compare against only two adapted versions of CNC-Net, while CNCGEN uses richer input representations and supervision. This makes it difficult to tell whether the gains come from the proposed method or from CNC-Net not being adapted effectively enough to the current setting.

Questions for authors

What is the main methodological contribution beyond integrating existing modules into a complete pipeline?

Why does CNC-Net leave substantial residual material even with ground-truth machining regions and depth information?

How do the authors handle multiple reasonable machining plans for the same part?

Reviewer_6TuyOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors introduce CNCGEN, a framework for learning machining process planning and tool path generation from B-rep CAD models. The systems tracks both the target and the metal remaining with every cut and adapts accordingly. During training, a frozen learned Surrogate Verifier is introduced to predict if any suggested cut leaves excess material behind or gouges the part, providing instant feedback to help improve the planner.

The authors also introduce a synthetically generated dataset of 50 thousand machining flow samples used for training and validation, and a held-out eval set of 800 CNC records. The authors evaluate CNCGEN on both synthetic and real benchmarks and demonstrate improvement over the CNC-Net baselines. The authors also conduct ablations that highlight state-conditioned tool path decoding’s importance.

Critical strengths
  1. The authors introduce the expansive CNCGEN-Dataset, a valuable dataset of about 50,000 geometrically verified synthetic machining flows and 800 real physical CNC machining flows (collected on Haas and Tormach CNC machines and validated by 50 CNC-domain experts). All machining flows in the synthetic subset densely map B-rep geometry to manufacturing objects, operation sequences, tool path parameterisations, stock states and corresponding verification outcomes.

  2. Ablation studies in the paper removing state-conditioned toolpaths, show that while sequence exact match drops only marginally from 0.829 to 0.826, Removal F1 drops significantly from 0.9016 to 0.3113, and overcut more than doubles from 0.0079 to 0.0169. This highlights that discrete operation planning alone is not enough to ensure successful material removal, and that tool motion must be conditioned on the intermediate stock evolution.

Critical weaknesses
  1. The baselines compared against are too weak, and don’t adequately validate what the method adds. The other learned baselines that the authors showcase is their own adaptation of CNC-Net, a self-supervised reconstruction method which appears to be unusually weak. Even when the exact removal region and depth are provided as an oracle, the method leaves about 60 percent of the intended material uncut. Additionally, the authors don’t include rule-based CAM relying on feature recognition with models like AAGNet and BRepGAT (which the authors cite). Since the reference flows are themselves rule-generated, this can potentially be a strong baseline.

  2. Evaluating on a 4.0mm occupancy grid is too coarse for the features described in the paper. Several features cover only a few voxels or just one (e.g. 3 mm holes and narrow chamfers). As such, IoU measurements are dominated by untouched stock (with the broken oracles baseline still scoring 0.927 despite leaving about 60% of the removal uncut), and Removal F1 becomes noisy with three different ablations being within similar values of 0.304 and 0.312. This coarse resolution may obscure important geometric errors, making it difficult to confidently interpret the ablation results.

  3. The paper does not report any measures for statistical variability. The authors only state that results are being averaged over three random seeds in Appendix C.1, however no STD, 95CIs, or other significance tests are provided anywhere in the manuscript. Considering many ablations have margins of less than 0.01, providing variances estimates are necessary to support conclusions.

Questions for authors
  1. Can the authors provide any measures of spread (stddev or 95CIs), across the three seeds conducted for all results in Tables 1-3?

  2. Why does CNC-Net + Oracle MRP leave ~60% of the intended material uncut (Residual 0.597 synthetic, 0.662 real) despite being given the exact removal region and depth?

  3. Could the authors also provide results for feature recognition followed by the same rule-based milling logic used to generate the synthetic flows?

1 comment

I want to thank the authors for their responses and added experiments. The revisions made satisfy the methodological concerns that I’ve raised, particularly regarding the addition of the AAGNet + generator rules baseline, evaluation-grid sensitivity, and variability across training seeds, and as such I’ll be increasing score.

Reviewer_wGKZOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies 3‑axis CNC process and toolpath generation from B‑Rep CAD models. It points out that correct operation sequences alone cannot guarantee good cutting results, since each cut changes stock material and influences later machining. The paper proposes an object‑state rollout idea and builds the CNCGEN dataset and its learning framework. The dataset includes 50 000 geometry‑checked synthetic samples and 800 real‑world CNC records only for testing. Each sample aligns B‑Rep geometry, machining steps, toolpaths and intermediate stock conditions. It supports joint supervision for discrete process decisions and continuous tool movement. The framework uses shared object memory to manage machining features. It adjusts operations and toolpaths based on past stock status. A frozen surrogate verifier provides differentiable geometric feedback in training. The model optimizes material‑removal quality, instead of only copying given process and toolpath labels. Compared with the CNC‑Net baseline, this method gets more complete material removal and less leftover material on both synthetic and real test data. Ablation tests prove that state‑aware toolpath generation and object memory are critical for final cutting quality. Its strengths are the practical object‑state view, well‑annotated dataset, real‑world test benchmark and careful experimental design. However, it only works for 3‑axis machining. It uses coarse 2.5D state representation, so small features and feature borders often leave residual material. Real‑world test samples are limited. Generated toolpaths need extra post‑processing for real machines. It only evaluates geometric results and ignores physical factors like cutting force and tool wear. Overall, this work offers valuable ideas and benchmarks for learning‑based CAM, but it cannot be directly used in full industrial production.

Critical strengths

Most prior work only focuses on predicting machining operation sequences. This paper points out that correct step sequences do not equal good real cutting results, because changed stock material affects later tool movements. The object‑state design connects each operation to its target feature and tracks evolving stock status. This captures key real‑world machining logic, and it forms the core idea of the whole work.

Critical weaknesses
  1. This framework only handles 3‑axis machining tasks. It cannot cover turning, multi‑axis machining or free‑form finishing. Besides, it uses a coarse 2.5D grid to track stock changes. Small features and feature boundary areas often keep unwanted residual material. These restrictions lower its practical value. Even with good test metrics, the method cannot solve many common real‑world CAM problems, and this hurts my overall assessment.
  2. All evaluations only check geometric cutting results. The paper does not measure physical factors such as cutting force, tool wear and tool vibration. Generated toolpaths also need machine‑specific post‑processing before real use. We cannot tell whether these predicted toolpaths are safe or suitable for actual machine execution. The gap between geometric simulation and real‑machine behavior weakens the real‑world validity of the presented results.
Questions for authors
  1. You use a coarse 2.5D grid to track stock state. Could you give quantitative numbers for how often residual material appears on narrow features or feature boundaries in your test set?
  2. Your current model only supports 3‑axis machining. In your dataset setup, what main barriers stop you from adding turning or multi‑axis machining samples for training and testing?
  3. All your evaluation focuses purely on geometric results. Do you have any test or metric to check whether your generated toolpaths avoid dangerous physical conditions like large cutting‑force spikes? If no, what concrete checks could you run to estimate real‑machine safety without full physical simulation?
  4. The 800 real‑world CNC records are only used for testing. What key differences exist between your synthetic training samples and these real records beyond feature shape? For example, do real‑world fixtures or tool limits create systematic performance gaps?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.