Pith. sign in

REVIEW 5 major objections 5 minor 61 references

3D-printing G-code can be generated directly from unoriented point clouds by predicting per-layer occupancy, region, and flow maps, bypassing mesh reconstruction entirely.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 02:27 UTC pith:QKY6P3TT

load-bearing objection Solid idea, real prints, but the Table 1 numbers hinge on an undefined sampling step - conditional accept, needs a precise evaluation protocol. the 5 major comments →

arxiv 2607.27729 v1 pith:QKY6P3TT submitted 2026-07-30 cs.CV

PrintAnything: Learning an Intermediate Representation for 3D printing G-code Generation

Sangmin Hong , Daniel Sungho Jung , Heewon Kim , Kyoung Mu Lee This is my paper
classification cs.CV
keywords 3D printingG-code generationpoint cloudsgeometric plan mapslice-wise projectiontoolpath synthesismesh-free fabricationinfill selection
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

PrintAnything claims that usable 3D-printing G-code can be produced directly from an unoriented point cloud, skipping the usual watertight-mesh requirement. To do this it predicts, per print layer, a "geometric plan" map—occupancy (where material goes), region (wall, infill, support, skirt), and flow (how much to extrude)—then compiles these maps into G-code with a hand-written slice converter. On a held-out set of 100K CAD–G-code pairs the reported accuracy is Chamfer distance 0.047, 3D F1 0.741, and slice-level 2D F1 0.677, all better than reconstructing a mesh first and then slicing it. If the claim holds, scanned or model-generated point clouds could be fabricated without surface repair, and toolpath generation would happen in about a third of a second per shape.

Core claim

On its own terms, the paper's discovery is that the watertight mesh is not a necessary intermediate for producing printable toolpaths. The authors define the G-plan map as the per-slice ground truth rasterized from existing slicer G-code: an occupancy mask M marking printed pixels, a region map R labeling each pixel as gap, wall, infill, support, or skirt, and a flow map Q carrying normalized extrusion length. A point-cloud encoder extracts a global shape feature, a per-slice decoder conditioned on slice height predicts M, R, Q at 256×256 resolution, an infill recommender inserts a hand-designed pattern into the infill region, and a deterministic compiler converts the result to G-code. Repor

What carries the argument

The load-bearing object is the Geometric plan (G-plan) map, a compact slice-wise 2D tensor with three channels: occupancy (printed vs empty), region (structural role such as wall, infill, support, skirt, gap), and flow (normalized extrusion amount). It matters because it is aligned with the layer-by-layer nature of fused-deposition printing by construction, so the network never has to reconstruct a surface; it only has to fill per-layer images. The companion mechanism is slice-wise projection of the point cloud onto the printer XY grid with adjacent slices as context, which lets the model reason about inter-layer continuity. A simple compiler turns the predicted maps into ordered perimeter a

Load-bearing premise

The pipeline assumes the input point cloud looks like 30,000 uniform samples of a watertight mesh normalized to [-1,1], because the supervision is rasterized from one slicer's G-code for exactly that distribution; if real scans differ in density, noise, orientation, or completeness, the predicted maps may not correspond to printable G-code.

What would settle it

Take the trained model and feed it a real scanner point cloud with non-uniform density and occlusion (or a point cloud sampled with a different number of points, e.g. 5k) from shapes in the same category, then slice the predicted G-plan maps and attempt to print. If the resulting prints show missing walls, detached infill, or fail, or if slice-level F1 drops dramatically compared to the 0.677 reported on the uniform-sample test set, the direct-to-G-code claim is confined to the training sampling distribution.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the claim holds, point clouds from commodity scanners or generative models can be turned into executable G-code in roughly 0.33 seconds, without mesh repair or manual slicing.
  • Slice-faithful toolpaths would avoid the failure propagation of mesh reconstruction: holes, inverted faces, and topological inconsistencies never enter the pipeline.
  • The per-shape infill recommender means material strength and cost can be traded off automatically, selecting among grid, cubic, honeycomb, and gyroid patterns.
  • Physical prints from the generated G-code include mechanical parts with thin structures, suggesting the maps retain the geometric detail mesh-based pipelines lose.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The ground truth is rasterized from a single slicer's G-code on watertight meshes, so the model is likely learning that slicer's conventions; a different slicer, nozzle width, or layer height would probably require retraining or map recalibration.
  • The reported robustness to noise and occlusion is qualitative; a natural stress test is to measure F1 and print success as point density, missing regions, and noise vary continuously.
  • Because the G-plan map is a generic per-slice image representation, the same recipe could extend to non-planar or multi-axis printing if the map channels were augmented with orientation or support constraints.
  • The paper acknowledges no explicit physics inductive bias; coupling the flow map to a melt-flow or strength model could push the approach beyond geometric imitation of one slicer.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The paper proposes PrintAnything, a supervised framework that takes unoriented point clouds as input and directly produces slice-wise 2D 'G-plan maps' — an occupancy map M, a region map R, and a flow map Q — which are then converted into executable G-code by a hand-written, rule-based slice-wise compiler. The stated contributions are a slice-aligned point projection strategy with multi-slice context, the G-plan representation itself, and an infill-pattern recommender. Experiments on the Slice-100K dataset report CD 0.047, F1_3D 0.741, and F1_2D 0.677, outperforming mesh-reconstruction-plus-slicer baselines (Poisson, DWG, MeshAnything followed by PrusaSlicer). Ablations show that each map component and multi-slice conditioning improve the reported metrics. Physical prints on a Bambu Lab X1-Carbon are shown as evidence of executability.

Significance. If the reported quantitative results are reproducible and the evaluation is made well-defined, the paper would offer a practical mesh-free route from point clouds to 3D-printing instructions for shapes similar to those in Slice-100K. The open-source release, the real physical prints, and the ablation evidence that all three G-plan components contribute are genuine strengths. However, the central quantitative claim currently rests on an under-specified evaluation protocol: Eq. (9) never defines how points are sampled from the 'predicted shape', the F1_3D threshold is not scale-consistent with the input normalization, and no error bars or statistical tests are reported. These gaps must be closed before the state-of-the-art claim can be accepted.

major comments (5)
  1. [Sec. 5.2, Eq. (9)] The metric definition is incomplete. 'Points sampled from the predicted shape' (P) is never defined, yet P is used in Eq. (9) to compute the headline CD and F1_3D numbers in Table 1. The model outputs per-slice 2D maps (Sec. 3.2) which are compiled to G-code (Sec. 3.4); no 3D point cloud or mesh is reconstructed. If P is sampled from the 2D occupancy masks at layer heights, then CD/F1_3D measures slice-raster agreement, not the geometry of a printable solid, and the comparison with mesh-based baselines (Poisson/DWG/MeshAnything + PrusaSlicer) is not like-for-like. Please specify precisely how P is obtained (e.g., by sweeping the toolpath volume derived from the generated G-code) and apply the same sampling to all compared methods.
  2. [Sec. 5.1, Sec. 5.2, Table 1] The numerical protocol is internally inconsistent. Inputs are normalized to [-1,1] (Sec. 5.1), but F1_3D uses a fixed distance threshold tau = 1 mm (Eq. 10). If coordinates are normalized, distances are dimensionless and 'mm' is meaningless; if coordinates are not normalized, the normalization statement is incomplete. Additionally, no error bars, confidence intervals, or significance tests are reported. The margins over DWG in Table 1 are small (CD 0.047 vs 0.062; F1_3D 0.741 vs 0.712), and a single random 9:1 split cannot establish that these differences are reliable. Report repeated-split or bootstrap statistics and clarify the physical scale handling.
  3. [Sec. 5.3] The comparison is not on equal footing. PrintAnything is trained on Slice-100K G-code labels, whereas the three baselines are generic mesh-reconstruction algorithms followed by an untuned PrusaSlicer pipeline. A task-specific supervised method can be expected to outperform generic baselines even if the G-plan representation contributes nothing. To support the claim that the representation is the source of improvement, add a supervised point-cloud-to-G-code baseline trained on the same split, or a simple voxelization-plus-PrusaSlicer baseline, and include an oracle bound obtained by rasterizing the GT G-code.
  4. [Sec. 5.6, Fig. 6, Supp. S6] The practical generalization claim is supported only qualitatively. The robustness experiments on incomplete and noisy point clouds (Fig. 5) and the ShapeNet results (Supp. Fig. S1) provide no quantitative CD, F1_2D, or print-success metrics, even though these inputs differ from the training distribution in exactly the factors (density, noise, completeness, scale) that matter for real scans. Likewise, the physical prints in Fig. 6 are selected examples without success rate, dimensional accuracy, or comparison against baseline G-code. Please add quantitative out-of-distribution evaluation and, ideally, a small quantitative physical-print study.
  5. [Sec. 5.5, Supp. S5] The infill recommender is trained and evaluated on the same hand-defined strength and cost proxies. Table 3 therefore demonstrates that the recommender optimizes these proxies, but it does not show that the selected infill patterns improve real structural strength or fabrication cost. The conclusion that the recommender leads to 'more efficient yet reliable fabrication outcomes' goes beyond the evidence. The appended Limitations (Supp. S11) already concedes that physics is not explicitly incorporated; this should be reflected in the claims and ideally supported by physical validation of at least a few recommended patterns.
minor comments (5)
  1. [Sec. 3.4, G-code example] The example G-code block is arithmetically inconsistent. Each 20 mm segment deposits volume 1.8 mm^3, corresponding to an extrusion length of about 0.75 mm. If E is absolute, the third move should be E2.25, not E1.05; if E is relative, all three moves should be E0.75. Please correct the example.
  2. [Table 4] The full model (M+R+Q) has a slightly worse CD (0.047) than M+R alone (0.046), while the text emphasizes only the improvements from adding Q. This small regression should be acknowledged or explained.
  3. [Supp. S7, Fig. S3] The region-map color mapping (gap/wall/infill/support/skirt) is described in the caption text of Supp. S7 but not directly in the figure caption. Adding a legend to the figure would improve readability.
  4. [Sec. 5.2] The F1_2D metric compares predicted slice occupancy against GT G-code rasterization. It is an imitation-fidelity metric relative to the training labels, not a physical print-quality metric. This should be stated explicitly when the metric is introduced.
  5. [Sec. 5.8] The statement that manual pressure tests showed 'the printed parts remained intact' is anecdotal and not quantified. If retained, it should be clearly labeled as informal observation, or replaced with a repeatable mechanical measurement.

Circularity Check

0 steps flagged

No significant circularity: PrintAnything is a supervised learning system whose central claim is supported by held-out evaluation and real-machine prints, not by a derivation that reduces to its inputs.

full rationale

PrintAnything is an empirical ML pipeline, not an analytic derivation, so the equation-level circularity patterns do not apply. The G-plan targets (M,R,Q) are rasterized from the same Slice-100K G-code used for F1_2D and GT geometry, but this is standard supervised learning with a 9:1 held-out split; the test metrics measure generalization to unseen shapes, and Fig. 6 reports actual prints on a Bambu Lab X1-Carbon, providing independent evidence of executability. The Chamfer/F1 definitions in Eqs. 9-11 leave the sampling of P underspecified, which is a reproducibility gap rather than a circular reduction: no equation shows that the predicted P is obtained from the fitted parameters by construction. The infill recommender (Sec. 3.3, Table 3) is trained and evaluated on the same strength/cost proxies, but the recommender is a surrogate fit to those proxies, not an identity; its output is selected from predicted S,C and then evaluated with the actual proxies, so the evaluation does not reduce to the training objective by definition. Self-citations (e.g., refs. 12,13,18,54,55) appear in related-work and limitations contexts and are not load-bearing for the central claim. The paper's own limitation (S11) notes the absence of physical inductive bias, but that is an acknowledged scope limitation, not a circular step. No specific circular step can be quoted with an exhibited reduction, so the appropriate finding is no significant circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 1 invented entities

The ledger shows the paper's contribution is a supervised representation-engineering result rather than a parameter-free derivation. The central method depends on many hand-set weights, a specific slicer dataset, and proxies for strength/cost. The only invented conceptual object is the G-plan map; no hidden physical entities are introduced.

free parameters (5)
  • Learned network weights (PTv3 encoder, U-Net decoder, FiLM conditioning, 1x1 heads, infill recommender regressors) = not enumerated; trained on Slice-100K
    All G-plan map predictions and infill choices depend on these trained weights; they are fit to rasterized GT G-code from Slice-100K and to the paper's own proxy labels.
  • Loss weights lambda_M, lambda_R, lambda_Q = 1.0, 1.0, 3e-5
    Hand-set in Sec. S4. The tiny flow weight means the claimed flow benefit in Table 4 depends on this specific balance.
  • Region class weights for weighted cross-entropy = [0.1, 3.0, 1.0, 2.0, 0.5]
    Hand-set in Sec. S4 to rebalance sparse classes; directly shapes the predicted region map R.
  • Architecture/hyper-parameters: voxel size 0.06, 256x256 maps, +/-1 slab context, 30k sampled points, 50 epochs, LR 2e-4 = as listed in Sec. 4/S3
    These manual choices determine the projection resolution, receptive field, and conditioning, and can change downstream toolpath quality.
  • Infill recommender proxy weights = Strength = 0.5*ratio + 0.3*connectivity + 0.2*balance; Cost = occ + 0.5*T_tot
    Hand-chosen in Sec. S5; defines the Pareto selection and Table 3 results. No mechanical validation ties these proxies to actual strength/cost.
axioms (4)
  • domain assumption GT G-code produced by PrusaSlicer on watertight meshes is a valid target for printable G-code.
    Used throughout to build GT G-plan maps (Sec. 3.1); any slicer bias (support placement, wall count, infill) is inherited by the model.
  • domain assumption Point clouds uniformly sampled from the target meshes represent the practical input regime.
    Training/test inputs on Slice-100K are 30k uniform surface samples of the same meshes whose G-code is used as supervision (Sec. 5.1); real scans are only shown qualitatively.
  • domain assumption The slice-wise XY projection with +/-1 slab context preserves enough information for per-layer print planning.
    Sec. 3.2/Table 2; vertical topology outside the slabs is carried only by a global feature vector, so cross-layer printability is largely implicit.
  • ad hoc to paper Infill proxies (material ratio, connectivity, directional balance) correlate with physical strength and fabrication cost.
    Sec. 3.3/S5; Table 3 claims strength/cost improvements but measures geometric proxies, not mechanical tests or measured cost.
invented entities (1)
  • G-plan map (M, R, Q) no independent evidence
    purpose: Intermediate 2D representation bridging point clouds and G-code; encodes occupancy, region semantics, and extrusion flow per slice.
    It is a paper-introduced representation, not an independently observed quantity. Its validity is established only through the paper's own downstream print/evaluation, so there is no external falsifiable handle beyond the reported results.

pith-pipeline@v1.3.0-daily-deepseek · 17513 in / 15378 out tokens · 162221 ms · 2026-08-01T02:27:34.193353+00:00 · methodology

0 comments
read the original abstract

Point clouds are one of the most fundamental and widely used 3D representations, serving as the most basic geometric representation of 3D shapes. Nevertheless, most existing 3D printing pipelines require a watertight mesh as input, preventing the direct use of point clouds for fabrication. A common workaround is to reconstruct meshes from point clouds; however, the resulting meshes often contain geometric artifacts, such as incorrect faces or topological inconsistencies, that are difficult to repair and may lead to printing failures. To overcome these limitations, we propose PrintAnything, a novel framework that learns to produce executable 3D printing G-code directly from 3D point clouds without requiring mesh reconstruction. To enable point clouds to serve as direct input for slice-wise toolpath generation, we introduce a slice-wise point projection strategy that transforms unstructured 3D point clouds into slice-aligned 2D representations consistent with layer-by-layer nature of fused deposition modeling in 3D printing. To eliminate mesh dependency and provide a unified representation that bridges point clouds and G-code, we propose Geometric plan (G-plan) map, a compact 2D representation composed of occupancy, region, and flow maps that encode the geometric and extrusion properties required for toolpath synthesis in 3D printing. As a result, our proposed method accurately generates printable G-code directly from point clouds, enabling a practical and fully mesh-free pipeline for 3D printing. The code is publicly available at \href{https://github.com/Sangminhong/PrintAnything}{https://github.com/Sangminhong/PrintAnything}.

Figures

Figures reproduced from arXiv: 2607.27729 by Daniel Sungho Jung, Heewon Kim, Kyoung Mu Lee, Sangmin Hong.

Figure 1
Figure 1. Figure 1: Motivated by Two Challenges: (a) No Point Cloud support for 3D Printing Slicers (b) Point Cloud to Mesh reconstruction is limited and error-prone. We present (c) PrintAnything converts point clouds into a G-plan Map and compiles slice-wise G-code for printing. naturally produced by LiDAR sensors [13], RGB-D cameras [12], and increas￾ingly by 3D generative models [8,15,18,23,34,41,43,46,60]. At the same tim… view at source ↗
Figure 2
Figure 2. Figure 2: The framework of PrintAnything. From input point cloud and slice indices, we encode them and pass to GPNet that predicts M, R, Q. The M, R, Q along with infill pattern dictionary are given to infill pattern generator to generate infilled M, R, Q. Lastly, we feed the infilled M, R, Q to layer-wise G-code generator to get predicted G-code which is directly converted to 3D print output. extrusion-based fabric… view at source ↗
Figure 3
Figure 3. Figure 3: GT G-plan map generation. We generate geometric plan (G-plan) map from a pair of G-code and 3D model. 3.1 GT Geometric plan (G-plan) map generation Given a pair consisting of a 3D mesh M and its corresponding sliced G-code G, we construct three geometric plan (G-plan) maps: an occupancy map M, a re￾gion map R, and a flow map Q. The sliced G-code G encodes the toolpath coordinates of the 3D printer together… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative Results on Slice-100K [16] dataset. mature slicer to produce G-code. Following this principle, we compare against strong mesh-conversion pipelines built on classical Poisson surface reconstruc￾tion [17] and recent learning-based methods, DWG [27] and MeshAnything [5]. For each baseline, we reconstruct a mesh from the input point cloud and feed the resulting mesh into PrusaSlicer [40] to generat… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative Results on incomplete and noisy point cloud input on [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of 3D printed output by Real Machine. [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

61 extracted references · 4 linked inside Pith

  1. [1]

    NeurIPS (2018) 29

    de Avila Belbute-Peres, F., Smith, K., Allen, K., Tenenbaum, J., Kolter, J.Z.: End- to-end differentiable physics for learning and control. NeurIPS (2018) 29

    2018

  2. [2]

    In: WACV

    Ben-Shabat, Y., Yu, X., Saleh, F., Campbell, D., Rodriguez-Opazo, C., Li, H., Gould, S.: The IKEA ASM Dataset: Understanding people assembling furniture through actions, objects and pose. In: WACV. pp. 847–859 (2021) 4

    2021

  3. [3]

    In: ICRA (2021) 4

    Bhatt, P.M., Kulkarni, A., Malhan, R.K., Gupta, S.K.: Optimizing part placement for improving accuracy of robot-based additive manufacturing. In: ICRA (2021) 4

    2021

  4. [4]

    arXiv preprint arXiv:1512.03012 (2015) 27

    Chang, A.X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., et al.: ShapeNet: An information-rich 3D model repository. arXiv preprint arXiv:1512.03012 (2015) 27

    Pith/arXiv arXiv 2015

  5. [5]

    In: ICLR 2, 10, 28, 29

    Chen, Y., He, T., Huang, D., Ye, W., Chen, S., Tang, J., Cai, Z., Yang, L., Yu, G., Lin, G., et al.: MeshAnything: Artist-created mesh generation with autoregressive Transformers. In: ICLR 2, 10, 28, 29

  6. [6]

    The Annals of Statistics (1976) 8, 25

    Collins, J.R.: Robust estimation of a location parameter in the presence of asym- metry. The Annals of Statistics (1976) 8, 25

    1976

  7. [7]

    ACM TOG (2018) 4

    Du, T., Inala, J.P., Pu, Y., Spielberg, A., Schulz, A., Rus, D., Solar-Lezama, A., Matusik, W.: InverseCSG: Automatic conversion of 3D models to CSG trees. ACM TOG (2018) 4

    2018

  8. [8]

    In: CVPR (2025) 2

    Du, Y., Zhao, Z., Su, S., Golluri, S., Zheng, H., Yao, R., Wang, C.: SuperPC: A single diffusion model for point cloud completion, upsampling, denoising, and colorization. In: CVPR (2025) 2

    2025

  9. [9]

    NeurIPS (2019) 29

    Greydanus, S., Dzamba, M., Yosinski, J.: Hamiltonian neural networks. NeurIPS (2019) 29

    2019

  10. [10]

    In: ICRA (2020) 3

    Gueners, D., Chanal, H., Bouzgarrou, B.: Stiffness optimization of a cable driven parallel robot for additive manufacturing. In: ICRA (2020) 3

    2020

  11. [11]

    ACM TOG (2020) 2

    Hanocka, R., Metzer, G., Giryes, R., Cohen-Or, D.: Point2Mesh: A self-prior for deformable meshes. ACM TOG (2020) 2

    2020

  12. [12]

    arXiv preprint arXiv:2512.13147 (2025) 2, 30

    Hong, S., Lee, S., Lee, K.M.: StarryGazer: Leveraging monocular depth estima- tion models for domain-agnostic single depth image completion. arXiv preprint arXiv:2512.13147 (2025) 2, 30

    arXiv 2025

  13. [13]

    In: CVPR (2023) 2, 30

    Hong, S., Yavartanoo, M., Neshatavar, R., Lee, K.M.: ACL-SPC: Adaptive closed- loop system for self-supervised point cloud completion. In: CVPR (2023) 2, 30

    2023

  14. [14]

    ACM TOG (2024) 4

    Huang, Y., Guo, Y., Su, R., Han, X., Ding, J., Zhang, T., Liu, T., Wang, W., Fang, G., Song, X., et al.: Learning based toolpath planner on diverse graphs for 3D printing. ACM TOG (2024) 4

    2024

  15. [15]

    In: CVPR (2024) 2

    Huang, Z., Johnson, J., Debnath, S., Rehg, J.M., Wu, C.Y.: PointInfinity: Resolution-invariant point diffusion models. In: CVPR (2024) 2

    2024

  16. [16]

    NeurIPS (2024) 4, 8, 9, 10, 11, 12, 13, 28, 31, 32

    Jignasu, A.N., Marshall, K., Mishra, A.K., Nerone Rillo, L., Ganapathysubrama- nian,B.,Balu,A.,Hegde,C.,Krishnamurthy,A.:Slice-100K:Amultimodaldataset for extrusion-based 3D printing. NeurIPS (2024) 4, 8, 9, 10, 11, 12, 13, 28, 31, 32

    2024

  17. [17]

    In: SGP (2006) 2, 10, 28, 29

    Kazhdan, M., Bolitho, M., Hoppe, H.: Poisson surface reconstruction. In: SGP (2006) 2, 10, 28, 29

    2006

  18. [18]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Kim, J., Yoo, J., Lee, J., Hong, S.: SetVAE: Learning hierarchical composition for generative modeling of set-structured data. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15059–15068 (2021) 2

    2021

  19. [19]

    In: CVPR (2019) 4 PrintAnything 17

    Koch, S., Matveev, A., Jiang, Z., Williams, F., Artemov, A., Burnaev, E., Alexa, M., Zorin, D., Panozzo, D.: ABC: A big CAD model dataset for geometric deep learning. In: CVPR (2019) 4 PrintAnything 17

    2019

  20. [20]

    ACM TOG (2015) 3

    Koyama, Y., Sueda, S., Steinhardt, E., Igarashi, T., Shamir, A., Matusik, W.: Au- toConnect: Computational design of 3D-printable connectors. ACM TOG (2015) 3

    2015

  21. [21]

    RA-L (2023) 4

    Ku, S., Choi, H., Kim, H.Y., Park, Y.L.: Automated sewing system enabled by machine vision for smart garment manufacturing. RA-L (2023) 4

    2023

  22. [22]

    gross motion

    Kubota, A., Iqbal, T., Shah, J.A., Riek, L.D.: Activity recognition in manufactur- ing: The roles of motion capture and sEMG+ inertial wearables in detecting fine vs. gross motion. In: ICRA (2019) 4

    2019

  23. [23]

    In: WACV (2025) 2

    Lee, J.J., Benes, B.: RGB2Point: 3D point cloud generation from single RGB im- ages. In: WACV (2025) 2

    2025

  24. [24]

    T-RO (2020) 3

    Li, J., Aubin-Fournier, P.L., Skonieczny, K.: SLAAM: Simultaneous localization and additive manufacturing. T-RO (2020) 3

    2020

  25. [25]

    In: ICCV (2025) 4

    Li, Y., Yuan, S., Wang, H., Li, Q., Liu, M., Xu, C., Shi, G., Zuo, W.: Triad: Empowering LMM-based anomaly detection with expert-guided region-of-interest tokenizer and manufacturing process. In: ICCV (2025) 4

    2025

  26. [26]

    ACM TOG (2024) 4

    Liu, T., Zhang, T., Chen, Y., Huang, Y., Wang, C.C.: Neural slicer for multi-axis 3D printing. ACM TOG (2024) 4

    2024

  27. [27]

    ACM TOG (2025) 2, 10, 28, 29

    Liu, W., Li, J., Chen, X., Hou, F., Xin, S., Wang, X., Wu, Z., Qian, C., He, Y.: Diffusing winding gradients (DWG): A parallel and scalable method for 3D reconstruction from unoriented point clouds. ACM TOG (2025) 2, 10, 28, 29

    2025

  28. [28]

    In: ICLR (2019) 8

    Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: ICLR (2019) 8

    2019

  29. [29]

    In: ICRA (2019) 4

    Malhan, R.K., Kabir, A.M., Shah, B., Gupta, S.K.: Identifying feasible work- piece placement with respect to redundant manipulator for complex manufacturing tasks. In: ICRA (2019) 4

    2019

  30. [30]

    In: ICRA (2023) 4

    Manyar, O.M., McNulty, Z., Nikolaidis, S., Gupta, S.K.: Inverse reinforcement learning framework for transferring task sequencing policies from humans to robots in manufacturing applications. In: ICRA (2023) 4

    2023

  31. [31]

    In: CVPR (2022) 29

    Mezghanni, M., Bodrito, T., Boulkenafed, M., Ovsjanikov, M.: Physical simulation layer for accurate 3D modeling. In: CVPR (2022) 29

    2022

  32. [32]

    In: 3DV (2016) 7, 24

    Milletari, F., Navab, N., Ahmadi, S.A.: V-Net: Fully convolutional neural networks for volumetric medical image segmentation. In: 3DV (2016) 7, 24

    2016

  33. [33]

    arXiv preprint arXiv:1312.5602 (2013) 4

    Mnih, V., Kavukcuoglu, K., Silver, D., Graves, A., Antonoglou, I., Wierstra, D., Riedmiller, M.: Playing Atari with deep reinforcement learning. arXiv preprint arXiv:1312.5602 (2013) 4

    Pith/arXiv arXiv 2013

  34. [34]

    Mo, S., Xie, E., Wu, Y., Chen, J., Nießner, M., Li, Z.: Fast training of diffusion transformerwithextrememaskingfor3Dpointcloudsgeneration.In:ECCV(2024) 2

    2024

  35. [35]

    In: ICRA (2021) 3, 4

    Munasinghe, N., Masangkay, J., Paul, G.: Temperature compensated 3D printed strain sensor for advanced manufacturing applications. In: ICRA (2021) 3, 4

    2021

  36. [36]

    RA-L (2020) 4

    Owan, P., Garbini, J., Devasia, S.: Faster confined space manufacturing teleoper- ation through dynamic autonomy with task dynamics imitation learning. RA-L (2020) 4

    2020

  37. [37]

    In: NeurIPS (2019) 8

    Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., Desmaison, A., Köpf, A., Yang, E., DeVito, Z., Raison, M., Tejani, A., Chilamkurthy, S., Steiner, B., Fang, L., Bai, J., Chintala, S.: PyTorch: An imperative style, high-performance deep learning library. In: NeurIPS (2019) 8

    2019

  38. [38]

    NeurIPS (2021) 2 18 S

    Peng, S., Jiang, C., Liao, Y., Niemeyer, M., Pollefeys, M., Geiger, A.: Shape as points: A differentiable poisson solver. NeurIPS (2021) 2 18 S. Hong et al

    2021

  39. [39]

    In: AAAI (2018) 6, 8, 24

    Perez, E., Strub, F., De Vries, H., Dumoulin, V., Courville, A.: FiLM: Visual reasoning with a general conditioning layer. In: AAAI (2018) 6, 8, 24

    2018

  40. [40]

    Prusa Research a.s.: PrusaSlicer.https://www.prusa3d.com/page/prusaslicer_ 424/(2026) 2, 8, 10

    2026

  41. [41]

    In: CVPR (2024) 2

    Ren, Z., Kim, M., Liu, F., Liu, X.: TIGER: Time-varying denoising model for 3D point cloud generation with diffusion process. In: CVPR (2024) 2

    2024

  42. [42]

    In: MICCAI (2015) 5, 8, 23

    Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolutional networks for biomed- ical image segmentation. In: MICCAI (2015) 5, 8, 23

    2015

  43. [43]

    In: CVPR (2022) 2

    Sanghi, A., Chu, H., Lambourne, J.G., Wang, Y., Cheng, C.Y., Fumero, M., Malek- shan, K.R.: CLIP-Forge: Towards zero-shot text-to-shape generation. In: CVPR (2022) 2

    2022

  44. [44]

    RA-L (2024) 4

    Schoonbeek, T.J., Balachandran, G., Onvlee, H., Houben, T., Hung, S.H., Kustra, J., de With, P.H., van der Sommen, F.: Supervised representation learning towards generalizable assembly state recognition. RA-L (2024) 4

    2024

  45. [45]

    In: ICRA (2025) 4

    Tong, X., Chang, Y., Zhao, Q., Yu, J., Wang, B., Lin, J., Lin, Y., Mai, X., Wang, H., Tao, Z., et al.: Component-aware unsupervised logical anomaly generation for industrial anomaly detection. In: ICRA (2025) 4

    2025

  46. [46]

    NeurIPS (2022) 2

    Vahdat, A., Williams, F., Gojcic, Z., Litany, O., Fidler, S., Kreis, K., et al.: LION: Latent point diffusion models for 3D shape generation. NeurIPS (2022) 2

    2022

  47. [47]

    In: ICRA (2025) 4

    Wang,Y.,Tang,B.,Gan,C.,Fox,D.,Mo,K.,Narang,Y.,Akinola,I.:MatchMaker: Automated asset generation for robotic assembly. In: ICRA (2025) 4

    2025

  48. [48]

    arXiv preprint arXiv:2511.20636 (2025) 4

    Wang, Z., Jadhav, Y., Pak, P., Farimani, A.B.: Image2Gcode: Image-to-G-code generation for additive manufacturing using diffusion-transformer model. arXiv preprint arXiv:2511.20636 (2025) 4

    arXiv 2025

  49. [49]

    Nature (2022) 29

    Wright, L.G., Onodera, T., Stein, M.M., Wang, T., Schachter, D.T., Hu, Z., McMa- hon, P.L.: Deep physical neural networks trained with backpropagation. Nature (2022) 29

    2022

  50. [50]

    ACM TOG (2016) 3

    Wu, R., Peng, H., Guimbretière, F., Marschner, S.: Printing arbitrary meshes with a 5DOF wireframe printer. ACM TOG (2016) 3

    2016

  51. [51]

    In: CVPR (2024) 5, 8, 23

    Wu, X., Jiang, L., Wang, P.S., Liu, Z., Liu, X., Qiao, Y., Ouyang, W., He, T., Zhao, H.: Point Transformer V3: Simpler faster stronger. In: CVPR (2024) 5, 8, 23

    2024

  52. [52]

    In: ICLR (2022) 4

    Yang, C., Wu, Z., Chee, J., De Sa, C., Udell, M.: How low can we go: Trading memory for error in low-precision training. In: ICLR (2022) 4

    2022

  53. [53]

    T-ASE (2023) 4

    Yang, H., Zhu, H., Li, J., Chen, J., Yin, Z.: Multi-category decomposition editing network for the accurate visual inspection of texture defects. T-ASE (2023) 4

    2023

  54. [54]

    In: CVPR (2024) 4

    Yavartanoo, M., Hong, S., Neshatavar, R., Lee, K.M.: CNC-Net: Self-supervised learning for CNC machining operations. In: CVPR (2024) 4

    2024

  55. [55]

    arXiv preprint arXiv:2411.06206 (2024) 4

    Yavartanoo, M., Hong, S., Neshatavar, R., Lee, K.M.: Text2CAD: Text to 3D CAD generation via technical drawings. arXiv preprint arXiv:2411.06206 (2024) 4

    Pith/arXiv arXiv 2024

  56. [56]

    In: WACV (2021) 4

    Zhang, G., Cui, K., Hung, T.Y., Lu, S.: Defect-GAN: High-fidelity defect synthesis for automated defect inspection. In: WACV (2021) 4

    2021

  57. [57]

    NeurIPS (2022) 4

    Zhang, M., Wang, X., Decardi-Nelson, B., Song, B., Zhang, A., Liu, J., Tao, S., Cheng, J., Liu, X., Yu, D., et al.: SMPL: Simulated industrial manufacturing and process control learning environments. NeurIPS (2022) 4

    2022

  58. [58]

    Zhang, T., Fang, G., Huang, Y., Dutta, N., Lefebvre, S., Kilic, Z.M., Wang, C.C.: S3-Slicer:Ageneralslicingframeworkformulti-axis3Dprinting.ACMTOG(2022) 4

    2022

  59. [59]

    NeurIPS (2023) 4 PrintAnything 19

    Zheng, S., Yang, H., Zhu, B., Yu, B., Wong, M.: LithoBench: Benchmarking AI computational lithography for semiconductor manufacturing. NeurIPS (2023) 4 PrintAnything 19

    2023

  60. [60]

    In: ECCV (2024) 2

    Zhou, C., Zhong, F., Hanji, P., Guo, Z., Fogarty, K., Sztrajman, A., Gao, H., Oztireli, C.: Frepolad: Frequency-rectified point latent diffusion for point cloud generation. In: ECCV (2024) 2

    2024

  61. [61]

    PrintAnything: Learning Geometric Plan Map for 3D Printing G-code Generation from Unoriented Point Clouds

    Zhou, Q., Jacobson, A.: Thingi10K: A dataset of 10,000 3D-printing models. arXiv preprint arXiv:1605.04797 (2016) 3 20 S. Hong et al. Supplementary Materialfor “PrintAnything: Learning Geometric Plan Map for 3D Printing G-code Generation from Unoriented Point Clouds" This supplementary material includes additional experiments, analyses, and results that c...

    Pith/arXiv arXiv 2016