Pith. sign in

REVIEW 4 major objections 6 minor 31 references

TransGraspNet claims that coupling boundary, surface, and physics consistency across perception, depth completion, and grasp scoring lets a robot reliably grasp and transport liquid-filled transparent labware, achieving high success in clut

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

2026-08-03 04:24 UTC pith:V2P55BR5

load-bearing objection A competent full-stack transparent-object manipulation system whose headline zero-spillage claim rests on a single unmeasured demo; worth reviewing but needs real repeats and a non-circular evaluation. the 4 major comments →

arxiv 2607.29567 v1 pith:V2P55BR5 submitted 2026-07-31 cs.RO

TransGraspNet: Physically and Geometrically Consistent Manipulation of Transparent Labware

Hailing Hu , Mingyi Zhu , Yiquan An , Yifei Tian , Tianyou Zuo , Lifeng Zhou This is my paper
classification cs.RO
keywords transparent object manipulationdepth completiongrasp planningwrench-space stabilityinstance segmentationliquid transportlaboratory automationconsistency
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.

TransGraspNet is a full-stack framework for robotic manipulation of transparent laboratory glassware, a safety-critical task where small geometric errors can tilt a grasp and spill hazardous liquids. The paper argues that the key problem is not any single perception module but cross-stage inconsistency: imperfect contours corrupt depth reconstruction, distorted surfaces corrupt normal estimation, and task-agnostic grasp scoring picks tilted or off-center grasps. To fix this, the system enforces three coupled principles: boundary consistency via an edge-supervised segmentation head, surface consistency via a gated depth-completion network that preserves curvature, and physics consistency via a grasp re-scoring step that aligns the grasp with the object’s centroid and principal axis and checks wrench-space stability. The central claim is that this integrated design yields reliable real-world operation: roughly 96% grasp success in simple scenes, 86% in clutter, and zero spillage during a 0.5 m/s liquid-transport test.

Core claim

The paper's central claim is that a transparent-object manipulation pipeline can achieve safety-grade performance if its stages are coupled by explicit geometric and physical consistency constraints, rather than optimized independently. Specifically, it asserts that (1) adding an edge-prediction branch to instance segmentation produces structurally reliable contours that serve as priors for depth completion; (2) an edge-guided attention gate and a masked geometric retention loss prevent cross-boundary depth bleeding and preserve surface normal fidelity, cutting normal error from 15.2° to 8.4°; and (3) re-scoring raw 6-DoF grasp candidates with a score that combines radial, angular, and centr

What carries the argument

The load-bearing mechanism is a staged consistency chain: (1) an Edge-Guided Boundary Consistency module that adds an E-CBAM attention and a dual-stream edge branch to Mask R-CNN-type detection, producing a mask and edge map that become explicit priors; (2) a Surface Consistency depth module (TDCNet backbone with an Edge-Guided Attention Gate and Masked Geometric Retention loss) that suppresses depth bleeding and preserves surface curvature, yielding a point cloud and normals; and (3) a Geometry–Physics Aware Grasp Refinement module that computes the object centroid and principal axis via PCA, then re-scores each grasp candidate with a weighted sum of radial alignment, angular matching, cent

Load-bearing premise

The physics-consistency grasp scores are computed from the object centroid and principal axis obtained by PCA on a depth-reconstructed point cloud whose error is substantial for small glassware; the paper does not show that these scores, and the resulting zero-spillage execution, are robust to realistic centroid and axis estimation error.

What would settle it

Take a real or simulated transparent vessel, run the pipeline to obtain a reconstructed point cloud, then deliberately perturb the reconstructed principal axis by 8° (the reported normal error) and the centroid by 10 mm; if the re-scored top-1 grasp changes to a tilted or off-center pose, or if a simulated transport with that perturbed score spills liquid, then the zero-spillage claim is not robust to the perception uncertainty the paper itself reports.

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

If this is right

  • If validated, this pipeline gives laboratory robots a practical way to handle transparent, liquid-containing vessels without spilling, which is a prerequisite for wider 'robot scientist' automation in chemistry and biology labs.
  • The explicit use of boundary priors in depth completion suggests that detection and reconstruction should be co-designed for transparent objects, challenging the common cascaded-module approach.
  • The physics-based grasp re-scoring step shows that task-level constraints (uprightness, centroid alignment, wrench stability) can be imposed as a post-processing filter on generic grasp candidates, making it a lightweight add-on to existing grasp detectors.
  • The reported normal-error reduction (15.2° to 8.4°) and corresponding grasp-orientation improvement indicate that surface normal fidelity is a strong predictor of downstream grasp quality for curved glassware.
  • The zero-spillage result links wrench-space scoring directly to dynamic outcomes, implying that stability-aware scoring can suppress inertial disturbances during high-speed transport.

Where Pith is reading between the lines

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

  • A critical unexamined link is the sensitivity of the PCA-derived centroid and principal axis to the reported depth error (RMSE 18.1 mm) and normal error (8.4°); on 50 ml-scale vessels these errors could bias the physics scores enough to change the selected grasp, so a perturbation analysis of the scoring step would clarify whether the zero-spillage guarantee is structurally stable.
  • The paper reports a single high-speed trajectory for liquid transport; extending the claim to arbitrary directions, varying liquid levels, or sloshing dynamics would require a dynamic model of fluid–vessel interaction, which the current wrench-space analysis does not include.
  • The boundary-consistency idea could generalize beyond labware to other specular or refractive objects (e.g., medical vials, food containers), where edge quality is the main bottleneck for downstream reconstruction.
  • Because the grasp re-scoring weights are fit offline to a small labeled set (200 grasps), the method's transferability to new grippers or object scales may require re-calibration; the paper does not address how the weights scale with object size.

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

4 major / 6 minor

Summary. The paper proposes TransGraspNet, a full perception-to-execution pipeline for grasping transparent laboratory glassware containing liquid. The system couples three consistency modules: edge-guided instance segmentation that supplies boundary priors, a depth-completion network with an edge-guided attention gate and masked geometric retention loss, and a grasp-scoring module that re-ranks 6-DoF candidates using radial/angular/centroid alignment terms and wrench-space robustness. The authors evaluate the components on the RobotSci-Glass dataset and ClearGrasp, and report real-robot grasp success rates of 96.0% in simple scenes and 86.0% in cluttered scenes, as well as 'zero spillage' in a dynamic liquid-transport demonstration.

Significance. If the claims hold, TransGraspNet addresses a real and safety-critical need in laboratory automation: transparent objects are notoriously difficult for RGB-D perception and for grasp selection, and the paper's core idea of coupling boundary quality, surface reconstruction, and physics-based scoring is sensible and potentially valuable. The paper also contributes a domain-specific dataset (RobotSci-Glass), clear experimental protocols for the static grasping trials, and task-level success metrics. However, the headline safety claim of zero-spillage dynamic transport is supported only by an anecdotal single demonstration, and part of the grasp-quality evaluation is close to tautological because the reported metrics are the same quantities that the proposed scoring terms explicitly optimize. The lack of sensitivity analysis linking depth reconstruction errors to the physics-based scores further weakens the current evidence for the central claim.

major comments (4)
  1. [§IV.F.3, Fig. 10] The 'zero spillage during high-speed liquid transport' claim is load-bearing for the paper's safety-oriented contribution, but the evidence is a single qualitative demonstration. No number of trials is reported, no repetitions, no pre/post mass measurement, no droplet detection, and no formal definition of spillage. The abstract and conclusion elevate this to 'reliable real-world operation' and 'zero-spillage dynamic stability.' A single successful run cannot establish reliability, particularly for a safety-critical claim. Please provide repeated trials (e.g., 20+), an objective spillage metric (mass loss or droplet sensor), and report the resulting statistics and confidence intervals.
  2. [§IV.D.3, Table III, Eqs. (10)–(12)] Table III reports that the proposed scoring achieves 3.8° angular error and 8.5 mm offset. But angle error is essentially the quantity optimized by S_angle in Eq. (11), and center offset is essentially the quantity optimized by S_centroid in Eq. (12). The comparison is therefore near-tautological: the method is being evaluated on its own objective rather than on an independent measure of task success. This does not validate the physics-consistency claim. Please add an evaluation that does not coincide with the optimized objectives, for example: (i) success rates on opaque objects with baseline vs. proposed scoring under the same full pipeline, or (ii) a perturbation analysis showing how the ranking changes when centroid/axis estimates are degraded by amounts consistent with Table II.
  3. [§IV.B.3, Table VI] The grasp score weights w0–w5 are calibrated offline using 200 labeled grasps by fitting a linear regression to maximize correlation with grasp success. The paper does not state whether these 200 grasps are disjoint from the real-robot trials in Table VI, nor whether any cross-validation was used. If the same data or the same experimental conditions informed the weights, the reported 96.0%/86.0% success rates may be optimistic. Please specify the train/test split for the weight calibration, the source of the 200 grasps, and whether the real-robot trials were independent of that calibration set.
  4. [§III.C, §IV.D.2, Table II] The physics scores in Eqs. (10)–(14) are computed from the object centroid and principal axis estimated by PCA on the reconstructed point cloud. Table II reports depth RMSE of 18.1 mm and surface normal error of 8.4° on RobotSci-Glass. For 50 ml-scale glassware these errors are a substantial fraction of the object size, and systematic biases such as flattened curvature or boundary bleeding could corrupt the principal-axis and centroid estimates, and therefore the 'physics-consistent' grasp selection and the upright-transport guarantee. No sensitivity analysis is provided. Please analyze how perturbations or realistic errors in the centroid and principal axis affect the final grasp ranking and the task-level outcomes, or otherwise justify that the scoring is robust to the observed reconstruction error.
minor comments (6)
  1. [Fig. 1 caption] Typo: 'TransGraspNet consist of' should be 'TransGraspNet consists of'.
  2. [Eq. (13)] The antipodal condition uses an indicator with condition 'π − θ_n < 2β'. Please define θ_n and β explicitly. As written, the condition is opaque.
  3. [Eq. (14)] The wrench-space metric S_Q = Radius(ConvexHull(Wrenches)) is not fully specified. Which wrenches are included, how the friction cone is discretized, and what coordinate frame is used? Without this, the term is not reproducible.
  4. [§II.B] The 'TDCNet' backbone is used throughout, but the reference list contains no explicit citation for TDCNet itself; the related-work discussion cites [19], [20] for gating mechanisms but not for the full backbone. Please add the appropriate citation or clarify the provenance.
  5. [§IV.D.3] The geometric-quality experiment is performed on 'opaque objects' but it is not stated how these objects relate to the transparent glassware in the real-robot trials. Please specify the object set and whether the same objects appear in Table III and Table VI.
  6. [§IV.F.3] Figure 10 has no axis labels or scale bar, and the liquid surface is only described qualitatively. Adding quantitative traces or at least a time-stamped sequence would make the dynamic test more informative.

Circularity Check

1 steps flagged

One local circularity: Table III's geometric-quality metrics are the optimizer's own angle/centroid objectives; central success-rate and depth claims remain independently grounded.

specific steps
  1. self definitional [Section IV.D.3, Table III; Eqs. (10)-(15)]
    "Angular matching: Sangle = exp(− (θapp − θideal)^2 / 2σa^2), (11) ... Centroid alignment: Scentroid = exp(− Δ^2_axial + Δ^2_lateral / 2σc^2). (12) ... Metrics. We compare Top-1 grasps to the ideal principal-axis frame using: (i) Angular Error (θ err) between approach vector and surface normal (smaller is more vertical); (ii) Center Offset (d of f) between grasp center and object centroid along height (smaller is more stable)."

    The final score (Eq. 15) is a weighted sum that directly includes S_angle and S_centroid, so selecting the top-1 grasp minimizes angular error to the PCA principal axis and centroid offset. Table III then reports exactly these two quantities (Angle Err, Offset) as evidence of geometric quality and credits the improvement to the physics-aware scoring. The reported angle/offset improvement is therefore the optimizer's own objective evaluated on itself, not an independent measure of grasp quality; the table demonstrates that the re-ranking optimizes its own scoring terms. The success-rate column and the depth/perception comparisons are separate measurements and are not circular.

full rationale

TransGraspNet's central derivation is mostly self-contained. Segmentation is benchmarked against TransLab/PointRend on ClearGrasp (Table IV), depth completion against ClearGrasp/NLSPN/TDCNet on public test data (Table V), and real-robot success rates in Table VI are separate executed trials after offline weight calibration. None of these reduce to the model's objectives by construction, and the paper does not rely on any load-bearing self-citation or imported uniqueness theorem. The one genuine circular element is localized: Table III's geometric quality metrics (Angle Err, Center Offset) are literally the quantities minimized by Eqs. 11 and 12 in the final score, so that particular comparison is partly the optimizer evaluating its own objective. The offline weight fitting to 200 success labels is supervised calibration and is not by itself circular; the 'zero spillage' dynamic test is under-reported (a single run with no repetitions or objective spillage measurement), but that is an evidence gap rather than a circularity and does not raise the circularity score. Overall, the safety-grade central claims rest on independent execution data, so the overall circularity is low-to-moderate.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

No new physical entities are introduced. The central claim rests on four modeling assumptions: PCA-proxy accuracy on imperfect reconstruction, candidate coverage by GraspNet, wrench-space validity on estimated normals, and transfer from opaque-coated supervision to transparent objects. Four fitted/hand-chosen parameter groups (score weights, kernel bandwidths, loss weights) control the headline numbers; the score weights are fit to labeled grasp outcomes.

free parameters (4)
  • grasp score weights w0-w5 = not reported
    Calibrated by linear regression on 200 grasps labeled success/failure to maximize correlation with success (Sec IV.B.3); values not given.
  • Gaussian bandwidths sigma_r, sigma_a, sigma_c = not reported
    Hand-chosen scale parameters in Eqs 10-12; values not reported.
  • edge loss weight lambda = not reported
    Eq 4; chosen without stated criterion.
  • MGR loss weights alpha_high, alpha_decay = not reported
    Eq 8; chosen without stated criterion.
axioms (4)
  • domain assumption PCA on the reconstructed point cloud yields a valid proxy for the true principal axis and centroid of glassware
    Sec III.C computes centroid and principal axis from imperfect depth (RMSE 18.1 mm, normal error 8.4 deg); no sensitivity analysis is given, and the assumption breaks for asymmetric vessels such as Erlenmeyer flasks with necks.
  • domain assumption GraspNet-1Billion's candidate generation covers the feasible grasp space for transparent labware
    Refinement only re-ranks GraspNet candidates (Sec IV.B.3), so good grasps missed by the generator cannot be recovered; the grasping model was trained on opaque household objects, not glass labware.
  • domain assumption Wrench-space robustness computed from estimated surface normals is a reliable stability proxy
    Eq 14 uses normals reconstructed with 8.4 deg mean error; normal error propagates into the wrench convex hull radius, and no robustness study is provided.
  • domain assumption Fine-tuning on opaque-coated depth ground truth transfers to genuinely transparent objects
    Sec IV.A.2 acquires near noise-free depth by blackening vessels; the paper assumes the network fine-tuned on these coated scenes generalizes to refractive, uncoated glass at test time (mitigated by ClearGrasp pretraining on synthetic transparent objects).

pith-pipeline@v1.3.0-daily-deepseek · 9316 in / 14411 out tokens · 151950 ms · 2026-08-03T04:24:46.153017+00:00 · methodology

0 comments
read the original abstract

Manipulating transparent laboratory glassware that contains liquid is inherently safety-critical: even small geometric errors can cause unstable grasps and hazardous spillage. Although recent progress has been made in transparent object perception and robotic grasping, most existing systems optimize detection, depth reconstruction, and grasp planning independently, which leads to cross-stage inconsistency imperfect boundaries induce depth bleeding, distorted surfaces corrupt normal estimation, and task agnostic grasp scoring yields tilted or off-center grasps that fail under dynamic motion. In this paper, we propose TransGraspNet, a geometry physics consistent framework that explicitly enforces consistency from perception to execution through three coupled principles: boundary consistency to produce structurally reliable object contours as downstream priors, surface consistency to preserve geometric fidelity and surface normal accuracy during depth reconstruction, and physics consistency to refine grasp selection with centroid alignment and wrench-space stability for upright and dynamically robust manipulation. We evaluate TransGraspNet on public benchmarks, a dedicated transparent glassware dataset, and a real robotic platform. The results show improved boundary quality and surface normal fidelity, and demonstrate strong task-level performance in cluttered transparent scenes. Most importantly, the proposed system achieves reliable real-world operation, including high grasp success rates in clutter and zero spillage during high speed liquid transport, highlighting the effectiveness of our method.

Figures

Figures reproduced from arXiv: 2607.29567 by Hailing Hu, Lifeng Zhou, Mingyi Zhu, Tianyou Zuo, Yifei Tian, Yiquan An.

Figure 1
Figure 1. Figure 1: Overall framework. TransGraspNet consist of perception, depth re [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Architecture of TransGraspNet-Det. The network employs E [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Architecture of TransGraspNet-Depth. The model integrates an [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Geometry-physics aware grasp refinement. Raw candidates are [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Qualitative comparison of depth completion. Error heatmaps show [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visualization of Top-k grasp ranking. Our geometry-physics aware re-ranking prioritizes structurally aligned and stable poses over the tilted or edge-biased grasps favored by GraspNet. TABLE IV COMPARISON OF INSTANCE SEGMENTATION METHODS ON CLEARGRASP Method APmask APbox Boundary F Mask R-CNN 76.9 78.4 52.1 PointRend 78.5 80.1 61.4 TransLab 79.8 81.2 58.6 Ours 78.2 79.5 65.1 As shown in Table III, the base… view at source ↗
Figure 9
Figure 9. Figure 9: Manipulation results in multi-object scenarios. The complete pick [PITH_FULL_IMAGE:figures/full_fig_p007_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Dynamic stability test in high-speed liquid transport. The robot [PITH_FULL_IMAGE:figures/full_fig_p007_10.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

31 extracted references · 2 linked inside Pith

  1. [1]

    Using specu- larities for recognition,

    Osadchy, Jacobs, and Ramamoorthi, “Using specu- larities for recognition,” inProceedings Ninth IEEE International Conference on Computer Vision, IEEE, 2003, pp. 1512–1519

    2003

  2. [2]

    Finding glass,

    K. McHenry, J. Ponce, and D. Forsyth, “Finding glass,” inProceedings of the 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recog- nition (CVPR), 2005, pp. 973–979.DOI:10.1109/ CVPR.2005.161

    2005

  3. [3]

    Transcut: Transparent object segmentation from a light-field image,

    Y . Xu, H. Nagahara, A. Shimada, and R.-i. Taniguchi, “Transcut: Transparent object segmentation from a light-field image,” inProceedings of the IEEE Interna- tional Conference on Computer Vision, 2015, pp. 3442– 3450

    2015

  4. [4]

    Deep polarization cues for transparent object segmentation,

    A. Kalra, V . Taamazyan, S. K. Rao, K. Venkataraman, R. Raskar, and A. Kadambi, “Deep polarization cues for transparent object segmentation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020, pp. 8602–8611.DOI:10. 1109/CVPR42600.2020.00863

    arXiv 2020

  5. [5]

    Transparent object detection using regions with convolutional neural network,

    P.-J. Lai and C.-S. Fuh, “Transparent object detection using regions with convolutional neural network,” in IPPR conference on computer vision, graphics, and image processing, vol. 2, 2015

    2015

  6. [6]

    Clear grasp: 3d shape estimation of transparent objects for manipulation,

    S. Sajjan et al., “Clear grasp: 3d shape estimation of transparent objects for manipulation,” in2020 IEEE international conference on robotics and automation (ICRA), IEEE, 2020, pp. 3634–3642

    2020

  7. [7]

    Segmenting transparent objects in the wild,

    E. Xie, W. Wang, W. Wang, M. Ding, C. Shen, and P. Luo, “Segmenting transparent objects in the wild,” inEuropean Conference on Computer Vision (ECCV), 2020

    2020

  8. [8]

    Segmenting transparent objects in the wild with transformer,

    E. Xie et al., “Segmenting transparent objects in the wild with transformer,” inProceedings of the Thirtieth International Joint Conference on Artificial Intelligence (IJCAI), 2021, pp. 1194–1200.DOI:10 . 24963 / ijcai.2021/165

    2021

  9. [9]

    Sparsity invariant cnns,

    J. Uhrig, N. Schneider, L. Schneider, U. Franke, T. Brox, and A. Geiger, “Sparsity invariant cnns,” in2017 international conference on 3D Vision (3DV), IEEE, 2017, pp. 11–20

    2017

  10. [10]

    Depth estimation via affinity learned with convolutional spatial propagation network,

    X. Cheng, P. Wang, and R. Yang, “Depth estimation via affinity learned with convolutional spatial propagation network,” inProceedings of the European conference on computer vision (ECCV), 2018, pp. 103–119

    2018

  11. [11]

    Non-local spatial propagation network for depth com- pletion,

    J. Park, K. Joo, Z. Hu, C.-K. Liu, and I. So Kweon, “Non-local spatial propagation network for depth com- pletion,” inEuropean conference on computer vision, Springer, 2020, pp. 120–136

    2020

  12. [12]

    Rgb-d local implicit function for depth completion of transparent objects,

    L. Zhu et al., “Rgb-d local implicit function for depth completion of transparent objects,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 4649–4658

    2021

  13. [13]

    Tran- scg: A large-scale real-world dataset for transparent object depth completion and a grasping baseline,

    H. Fang, H.-S. Fang, S. Xu, and C. Lu, “Tran- scg: A large-scale real-world dataset for transparent object depth completion and a grasping baseline,” IEEE Robotics and Automation Letters, vol. 7, no. 3, pp. 7383–7390, 2022

    2022

  14. [14]

    Sparse-to-dense: Depth pre- diction from sparse depth samples and a single image,

    F. Ma and S. Karaman, “Sparse-to-dense: Depth pre- diction from sparse depth samples and a single image,” in2018 IEEE international conference on robotics and automation (ICRA), IEEE, 2018, pp. 4796–4803

    2018

  15. [15]

    Learn- ing guided convolutional network for depth comple- tion,

    J. Tang, F.-P. Tian, W. Feng, J. Li, and P. Tan, “Learn- ing guided convolutional network for depth comple- tion,”IEEE Transactions on Image Processing, vol. 30, pp. 1116–1129, 2020

    2020

  16. [16]

    Deeplidar: Deep surface normal guided depth prediction for outdoor scene from sparse lidar data and single color image,

    J. Qiu et al., “Deeplidar: Deep surface normal guided depth prediction for outdoor scene from sparse lidar data and single color image,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 3313–3322

    2019

  17. [17]

    Revisiting single image depth estimation: Toward higher resolu- tion maps with accurate object boundaries,

    J. Hu, M. Ozay, Y . Zhang, and T. Okatani, “Revisiting single image depth estimation: Toward higher resolu- tion maps with accurate object boundaries,” in2019 IEEE winter conference on applications of computer vision (WACV), IEEE, 2019, pp. 1043–1051

    2019

  18. [18]

    Sparse and dense data with cnns: Depth completion and semantic segmentation,

    M. Jaritz, R. De Charette, E. Wirbel, X. Perrotton, and F. Nashashibi, “Sparse and dense data with cnns: Depth completion and semantic segmentation,” in2018 International Conference on 3D Vision (3DV), IEEE, 2018, pp. 52–60

    2018

  19. [19]

    Squeeze-and-excitation networks,

    J. Hu, L. Shen, and G. Sun, “Squeeze-and-excitation networks,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 7132–7141

    2018

  20. [20]

    A multi-scale guided cascade hourglass network for depth completion,

    A. Li, Z. Yuan, Y . Ling, W. Chi, C. Zhang, et al., “A multi-scale guided cascade hourglass network for depth completion,” inProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2020, pp. 32–40

    2020

  21. [21]

    Robotic grasping and con- tact: A review,

    A. Bicchi and V . Kumar, “Robotic grasping and con- tact: A review,” inProceedings 2000 ICRA. Millen- nium conference. IEEE international conference on robotics and automation. Symposia proceedings (Cat. No. 00CH37065), IEEE, vol. 1, 2000, pp. 348–353

    2000

  22. [22]

    Deep learning for detecting robotic grasps,

    I. Lenz, H. Lee, and A. Saxena, “Deep learning for detecting robotic grasps,”The International Journal of Robotics Research, vol. 34, no. 4-5, pp. 705–724, 2015

    2015

  23. [23]

    Real-time grasp detection using convolutional neural networks,

    J. Redmon and A. Angelova, “Real-time grasp detection using convolutional neural networks,” in2015 IEEE international conference on robotics and automation (ICRA), IEEE, 2015, pp. 1316–1322

    2015

  24. [24]

    Closing the loop for robotic grasping: A real-time, generative grasp synthesis approach,

    D. Morrison, P. Corke, and J. Leitner, “Closing the loop for robotic grasping: A real-time, generative grasp synthesis approach,”arXiv preprint arXiv:1804.05172, 2018

    Pith/arXiv arXiv 2018

  25. [25]

    Grasp pose detection in point clouds,

    A. Ten Pas, M. Gualtieri, K. Saenko, and R. Platt, “Grasp pose detection in point clouds,”The Interna- tional Journal of Robotics Research, vol. 36, no. 13-14, pp. 1455–1473, 2017

    2017

  26. [26]

    Pointnetgpd: Detecting grasp configu- rations from point sets,

    H. Liang et al., “Pointnetgpd: Detecting grasp configu- rations from point sets,” in2019 International Confer- ence on Robotics and Automation (ICRA), IEEE, 2019, pp. 3629–3635

    2019

  27. [27]

    Dex-net 2.0: Deep learning to plan robust grasps with synthetic point clouds and analytic grasp metrics,

    J. Mahler et al., “Dex-net 2.0: Deep learning to plan robust grasps with synthetic point clouds and analytic grasp metrics,”arXiv preprint arXiv:1703.09312, 2017

    Pith/arXiv arXiv 2017

  28. [28]

    Graspnet- 1billion: A large-scale benchmark for general object grasping,

    H.-S. Fang, C. Wang, M. Gou, and C. Lu, “Graspnet- 1billion: A large-scale benchmark for general object grasping,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 11 444–11 453

    2020

  29. [29]

    Contact-graspnet: Efficient 6-dof grasp generation in cluttered scenes,

    M. Sundermeyer, A. Mousavian, R. Triebel, and D. Fox, “Contact-graspnet: Efficient 6-dof grasp generation in cluttered scenes,” in2021 IEEE International Confer- ence on Robotics and Automation (ICRA), IEEE, 2021, pp. 13 438–13 444

    2021

  30. [30]

    Trans- grasp: A multi-scale hierarchical point transformer for 7-dof grasp detection,

    Z. Liu, Z. Chen, S. Xie, and W.-.-S. Zheng, “Trans- grasp: A multi-scale hierarchical point transformer for 7-dof grasp detection,” in2022 International Confer- ence on Robotics and Automation (ICRA), IEEE, 2022, pp. 1533–1539

    2022

  31. [31]

    Graspdiff: Grasping generation for hand-object interaction with multimodal guided diffusion,

    B. Zuo, Z. Zhao, W. Sun, X. Yuan, Z. Yu, and Y . Wang, “Graspdiff: Grasping generation for hand-object interaction with multimodal guided diffusion,”IEEE Transactions on Visualization and Computer Graphics, 2024

    2024