TD-TOG Dataset: Benchmarking Zero-Shot and One-Shot Task-Oriented Grasping for Object Generalization

Andrew Starkey; Bruno Yun; Dewei Yi; Jamie Grech; Pascal Mei{\ss}ner; Valerija Holomjova

REVIEW 4 major objections 7 minor 6 references

TD-TOG Dataset: Benchmarking Zero-Shot and One-Shot Task-Oriented Grasping for Object Generalization

T0 review · 4 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A new real-world dataset and a zero-shot framework let robots grasp task-relevant parts of objects they have never seen, reaching 68.9% task-oriented grasp accuracy in multi-object scenes.

desk verdict Solid dataset contribution, but the 68.9% generalization claim is instance-level, not novel-instance generalization. read the letter →

arxiv 2506.05576 v1 pith:VGM3D3FW submitted 2025-06-05 cs.RO

Valerija Holomjova , Jamie Grech , Dewei Yi , Bruno Yun , Andrew Starkey , Pascal Mei{ss}ner This is my paper

classification cs.RO

keywords task-orientedgraspingzero-shotlearningone-shotroboticbenchmarkRGB-Ddatasetaffordancerecognitionobjectgeneralizationfoundationmodels

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper introduces TD-TOG, a hand-annotated real-world dataset of 1,449 top-down RGB-D scenes containing 30 object categories and 120 subcategories, with object masks, affordance masks, and planar grasp rectangles for every object. It also proposes Binary-TOG, a modular grasp-selection framework that uses zero-shot object recognition from a text prompt, one-shot affordance recognition from a reference image, and a grasp detector, so that it can pick task-relevant grasps on objects it was never trained on. The central empirical claim is that Binary-TOG reaches an average task-oriented grasp accuracy of 68.9% in multi-object scenes on the dataset, and that a standard framework trained on TD-TOG reaches 90.9% on known objects and 65.9% on new subcategories. The paper argues that these results make TD-TOG a usable benchmark and that the comparison clarifies when one-shot versus zero-shot generalization is preferable.

What carries the argument

The load-bearing components are the TD-TOG dataset itself and the modular Binary-TOG pipeline. TD-TOG provides 1,449 real-world RGB-D scenes with hand-annotated object masks, affordance masks, and oriented grasp rectangles, plus a separate subcategory test split that forces recognition of fine-grained object variants. Binary-TOG's pipeline consists of four modules: a zero-shot segmenter (SAM) filtered by a size/subset algorithm (SSF) that removes noise and duplicate masks; a zero-shot recognizer (CLIP) that matches masked scene crops to a text description of the target object; a one-shot affordance module (AffCorrs) that transfers a task-relevant region from a reference object; and a grasp detector (S2A-Net) whose candidates are filtered by the predicted affordance region. The modularity is what lets the authors attribute failures to specific components and compare zero-shot against one-shot generalization fairly.

What would settle it

Take a random sample of, say, 50 TD-TOG scenes, have two independent annotators re-label all objects, affordances, and grasps from scratch, and measure IoU and grasp-angle agreement. If the average mask IoU or grasp agreement falls well below the thresholds used to score the frameworks (IoU > 25% and angle within 30°), the reported accuracies such as 90.9% and 68.9% would not be reproducible against independent ground truth.

Watch

Extended reading notes

Core claim

The central claim is that TD-TOG provides a real-world, multi-object, hand-annotated TOG dataset that supplies all three annotation types needed for training (object masks, affordance masks, grasp rectangles), and that Binary-TOG, by combining zero-shot object recognition (CLIP), one-shot affordance transfer (AffCorrs), and a trained grasp detector (S2A-Net) behind a filtering step, can select task-relevant grasps on novel objects without retraining. On the TD-TOG category split, Binary-TOG attains an average task-oriented grasp accuracy of 68.9% in multi-object scenes, while a standard framework trained on TD-TOG attains 90.9% on known objects and 65.9% on new subcategories. In real-world trials with a robotic arm, Binary-TOG succeeded in 64.8% of KC-USC grasps and 52.7% of UC-USC grasps.

Load-bearing premise

The hand-annotated masks, affordances, and grasp rectangles in TD-TOG are accurate and consistent enough to serve as ground truth, yet the paper reports no inter-annotator agreement or annotation error rate.

Editorial extensions

If this is right

TD-TOG can serve as a standard training and evaluation set for task-oriented grasping, since it is real-world, multi-object, and contains all annotations required by typical TOG pipelines.
A framework that uses only a textual description for object recognition can generalize to new object categories without retraining, at the cost of lower fine-grained subcategory accuracy than visual one-shot references.
The new subcategory split reveals a previously untested capability—distinguishing near-identical object variants—and both tested frameworks lose accuracy there, so future TOG research should target this gap.
Choosing between zero-shot and one-shot generalization in TOG can be guided by the trade-off this paper measures: annotation effort versus recognition accuracy on visually subtle distinctions.
Real-world success rates (64.8% and 52.7% for Binary-TOG on new subcategories and new categories) give an estimate of how dataset accuracy transfers to physical execution.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A direct extension of the paper's modular design: swapping the zero-shot recognizer for a more recent vision-language model might raise subcategory accuracy without changing the rest of the pipeline, since the recognition module is the identified bottleneck.
The subcategory split could be reused as a cheap probe for the fine-grained visual grounding of any foundation model, independent of grasping, because it isolates recognition difficulty from grasp difficulty.
The paper's annotation pre-labeling algorithms, if released with the dataset, could substantially cut the cost of building similar TOG datasets for new object sets, but their output quality is only as good as the reference masks they start from.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The dataset is worth having. Real-world multi-object scenes, hand-annotated masks, affordances, grasps, plus a subcategory split that nobody else has. The authors also run physical robot trials, which is more than most dataset papers bother with. But the headline number—68.9% average TOG accuracy in multi-object scenes—does less work than the abstract says. The reference set contains one image per subcategory, and the test and real-world sets reuse those exact subcategories. So the one-shot affordance transfer (AffCorrs) is matching between two images of the same physical object in different scenes. That's instance-level viewpoint/background generalization, not generalization to a never-seen object of the same category. The subcategory split does test fine-grained discrimination, and that is genuinely new, but the 'object generalization' framing is overstated.

The dataset's training-test split is otherwise reasonable: Standard-TOG trained on TD-TOG hits 90.8% on known subcategories and 59.1% on novel subcategories, which suggests the annotations are usable. The SSF improvement over raw SAM is large and clearly presented. The comparison between one-shot and zero-shot recognition is useful even if the winner—OS-TOG—is not surprising.

Soft spots beyond the instance issue: no error bars or significance tests anywhere, and the partial-success weighting in real-world trials is underspecified (what makes a success 0.5 vs 1?). No inter-annotator agreement or annotation error rate, which matters because the masks are the ground truth. The dataset and code are 'to be released', so none of the numbers can be checked right now. The typo in the category split description ('not present in the testing set' should be 'training set') is minor but suggests a hasty final pass.

The stress-test note's reading holds up: if the reference and test objects are the same physical instances, the central generalization claim needs rewording. But that is a fixable presentation issue, not a fatal design flaw. The dataset itself is a real contribution and the subcategory challenge is a good idea.

For a venue that takes datasets seriously, this deserves peer review. Require the dataset and code as supplementary material, ask for confidence intervals and annotation quality metrics, and ask the authors to clarify the physical-instance overlap between references and test scenes. It's a conditional accept with revisions, not a desk reject.

Referee Report

4 major / 7 minor

Summary. The paper introduces TD-TOG, a real-world top-down RGB-D dataset of 1,449 scenes spanning 30 object categories, 120 subcategories, and 16 affordances, with hand-annotated object masks, affordance masks, and planar grasp rectangles. It also proposes Binary-TOG, a modular TOG framework that combines zero-shot object segmentation (SAM), zero-shot object recognition from textual prompts (CLIP), one-shot affordance recognition (AffCorrs), and a trained grasp detector (S2A-Net). The authors compare Binary-TOG with their earlier one-shot framework OS-TOG and with a Standard-TOG baseline on the dataset's category and subcategory splits, and they report real-world robot experiments. The central claims are that TD-TOG enables training and benchmarking of TOG solutions, that Binary-TOG achieves an average task-oriented grasp accuracy of 68.9% in multi-object scenes without retraining, and that the comparison illuminates trade-offs between zero-shot and one-shot object generalization.

Significance. If the claims hold, TD-TOG would be a valuable community resource: it is a real-world, multi-object, manually annotated TOG dataset with object masks, affordance masks, and grasps, addressing a genuine gap identified in Table I. The subcategory-discrimination test set is a novel and useful evaluation idea. The modular Binary-TOG framework is also a strength because it enables per-module error analysis and a cleaner comparison of zero-shot versus one-shot recognition than end-to-end baselines. The paper ships dataset, code, and models, and the comparative analysis is directly useful for practitioners. However, these contributions are currently undermined by a load-bearing ambiguity about whether the test objects are physically distinct instances from the reference objects, and by the absence of annotation-quality metrics, confidence intervals, and significance tests.

major comments (4)

[Section III-A and Section IV-B/IV-C] The object-generalization claims appear to be instance-level re-identification rather than generalization to previously unseen object instances. Each subcategory has exactly one reference image in the reference set (Section III-A), and the category-split test subsplits KC-KSC, KC-USC, and UC-USC reuse exactly the subcategories _01 and _04 of the known and unknown categories. At inference, OS-TOG and Binary-TOG retrieve the reference entry for the target subcategory and pass it to the one-shot affordance module (AffCorrs). Unless the authors explicitly state and ensure that test scenes contain physically different instances of the same subcategory, the reference and test images show the same physical object, differing only in pose, lighting, and background. Under that reading, the reported 68.9% average TOG accuracy (Table V, Section V-C) and the OS-TOG/Binary-TOG comparison measure viewpoint and scene robustness for known instances, not the 'previously unseen objects' claimed in the abstract and contribution 2. This ambiguity is load-bearing for the central contribution, and the paper never addresses it.
[Section III-C] The dataset's central quality claim - 'hand-annotated' masks, affordances, and grasps that are 'highly detailed' and free of the artifacts found in prior datasets - is not supported by any quantitative annotation-quality metric. No inter-annotator agreement, annotation error rate, or consistency check is reported. Because every framework accuracy in Tables IV, V, and X is computed against these annotations, systematic errors or inconsistencies in the masks or grasps would directly affect all reported figures, including the 90.9% and 68.9% headline numbers. Please add an annotation-quality analysis, such as a second-annotator IoU study on a subset, or at minimum clearly report the annotation review protocol's outcomes.
[Section V-B and Tables IV-V, X] No error bars, confidence intervals, or significance tests are reported for the dataset experiments or the real-world trials. The comparative conclusions - for example, that OS-TOG is 'more robust' in multi-object scenes (Section V-C) or that Binary-TOG and OS-TOG are 'relatively similar' in real-world KC-USC trials (Section V-D) - rest on differences such as 69.5% vs. 70.8% (Table V) and 64.8% vs. 65.3% (Table X). Given the modest number of scenes and trials, these differences may be within noise. The paper should report confidence intervals or perform significance testing before drawing comparative conclusions.
[Section V-B2] The partial-success rule for real-world experiments is underspecified. The text states that 'success can be partial (i.e. be 0.5 instead of 1) in cases where the task succeeded but the prediction was not ideal (e.g. the object segmentation only segmenting half of the object but managing to predict and execute the correct task-oriented grasp)', but it does not formally define when a trial receives 0.5, how partial successes are weighted in the reported averages, or what fraction of trials were partial across frameworks. Since all success rates in Table X are aggregates of these scores, the rule can materially change the reported numbers. Please specify the scoring function exactly and report the distribution of full, partial, and failed trials.

minor comments (7)

[Section IV-C] The formula for OS-TOG object recognition uses argmax over the L2 distance, but the text says the module selects the feature with the 'minimum' L2 distance; the equation should use argmin to match the stated behavior.
[Table XII] The header of Table XII reads 'Hybrid-TOG' in several places, but the framework is called Binary-TOG throughout the paper; the table header should be corrected.
[Section IV-B3, Algorithm 2] The number of rotations n_rots used in the Affordance Alignment module is never specified; please report the value and, ideally, a brief sensitivity check.
[Section V-A1] The SSF thresholds (min_area=400px2, max_area=50,000px2, tau=0.75) and the CLIP scaling factor of 100 are stated as empirically chosen, but no sensitivity analysis is provided; a short discussion of how sensitive the results are to these values would strengthen the reproducibility.
[Tables VI and VII] The captions of Tables VI and VII say 'category split and KC-USC subsplit' and 'category split and UC-USC subsplit', but the content appears to be per-object TOG accuracies; please clarify the exact data source and define what 'category split' means in this context to avoid confusion.
[Section V-D] The sentence 'Table XII showed that Binary-TOG struggled most in the UC-USC subsplit...' should be cross-checked: the referenced table appears to be Table XII in the current manuscript, but the text seems to intend the table showing UC-USC real-world results; please verify the reference.
[Section III-A and Table I] Table I marks TD-TOG as 'Publicly Available', but the abstract says 'will be released upon publication'; please reconcile this wording and ensure the release links are stable.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found: the headline accuracies are empirical benchmark results with external models and explicit train/test splits, not predictions forced by construction.

full rationale

I find no circular derivation chain in this paper. The central claims are empirical: TD-TOG is a new dataset, and the reported accuracies (90.9%, 68.9%, etc.) are computed by running frameworks on held-out test splits of that dataset. The grasp detector S2A-Net and the Standard-TOG sub-models are trained on the TD-TOG training split and evaluated on the category and subcategory test splits, which is the standard benchmark protocol rather than a fitted parameter being renamed as a prediction. Binary-TOG's recognition and affordance modules rely on external foundation models (CLIP, SAM, AffCorrs) that are not trained on TD-TOG; the paper explicitly notes that CLIP was not trained on TD-TOG. The one-shot affordance transfer uses a knowledge-base reference image of the target object, but providing that reference is the definition of the one-shot setting, not a hidden fit to the test quantity. The comparison with OS-TOG (Holomjova et al., 2023) is a legitimate baseline comparison with prior work, and no load-bearing claim rests on a self-citation or on a uniqueness theorem imported from the authors. Two caveats exist but are not circularity: the paper does not report inter-annotator agreement for its hand annotations, and it does not state whether test objects are physically distinct from the reference images for each subcategory. If they are the same instances, then 'object generalization' is partly instance re-identification, but that would be an evaluation-validity concern, not a reduction of the derivation to its own inputs. The reported numbers remain empirical measurements on held-out scenes with external models.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The paper does not introduce new theoretical entities. It makes empirical contributions (dataset, framework, challenge) and the few hand-set parameters are standard engineering choices rather than free parameters of a model. The main assumptions are domain-level: the correctness of the task rules and the representativeness of the top-down setup.

free parameters (5)

SSF min_area = 400 px^2
Hand-set threshold to remove small noise artifacts in SAM segmentations.
SSF max_area = 50,000 px^2
Hand-set threshold to remove large background regions in SAM segmentations.
SSF overlap threshold tau = 0.75
Hand-set threshold to remove redundant subset masks in SSF.
CLIP similarity scaling factor = 100
Multiplying cosine similarity by 100 to highlight score differences; an arbitrary constant.
n_rots in AA = not specified
The number of rotations in the affordance alignment algorithm is not reported, leaving an implementation detail open.

assumptions (3)

domain assumption The task-to-affordance rules in Table II correctly capture the regions needed for each task
The evaluation of TOG accuracy depends on these rules being correct; they are defined by the authors without external validation.
domain assumption Top-down RGB-D scenes of objects on a table are representative of relevant real-world TOG scenarios
The dataset only covers top-down views and objects placed in a fixed green region; generalizability to cluttered or multi-angle scenes is not established.
domain assumption The pre-trained models (SAM, CLIP, AffCorrs) provide valid zero-shot and one-shot capabilities without further training on TD-TOG
The framework's performance depends on these foundation models generalizing to the dataset's objects; preliminary trials are mentioned but not quantified.

how reviews work

0 comments

Cite this review

Pith. "Pith review of TD-TOG Dataset: Benchmarking Zero-Shot and One-Shot Task-Oriented Grasping for Object Generalization." pith.science (2026). https://pith.science/paper/VGM3D3FW

@misc{pith2026250605576,
  author       = {Pith},
  title        = {Pith review of: TD-TOG Dataset: Benchmarking Zero-Shot and One-Shot Task-Oriented Grasping for Object Generalization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VGM3D3FW}},
  note         = {Machine review of arXiv:2506.05576}
}

read the original abstract

Task-oriented grasping (TOG) is an essential preliminary step for robotic task execution, which involves predicting grasps on regions of target objects that facilitate intended tasks. Existing literature reveals there is a limited availability of TOG datasets for training and benchmarking despite large demand, which are often synthetic or have artifacts in mask annotations that hinder model performance. Moreover, TOG solutions often require affordance masks, grasps, and object masks for training, however, existing datasets typically provide only a subset of these annotations. To address these limitations, we introduce the Top-down Task-oriented Grasping (TD-TOG) dataset, designed to train and evaluate TOG solutions. TD-TOG comprises 1,449 real-world RGB-D scenes including 30 object categories and 120 subcategories, with hand-annotated object masks, affordances, and planar rectangular grasps. It also features a test set for a novel challenge that assesses a TOG solution's ability to distinguish between object subcategories. To contribute to the demand for TOG solutions that can adapt and manipulate previously unseen objects without re-training, we propose a novel TOG framework, Binary-TOG. Binary-TOG uses zero-shot for object recognition, and one-shot learning for affordance recognition. Zero-shot learning enables Binary-TOG to identify objects in multi-object scenes through textual prompts, eliminating the need for visual references. In multi-object settings, Binary-TOG achieves an average task-oriented grasp accuracy of 68.9%. Lastly, this paper contributes a comparative analysis between one-shot and zero-shot learning for object generalization in TOG to be used in the development of future TOG solutions.

Figures

Figures reproduced from arXiv: 2506.05576 by the authors.

**Figure 2.** An example scene from the TD-TOG training set annotated [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** A preview of 15 object categories (e.g. cable, clip) and 60 object subcategories (e.g. cable_01, cable_02) from the TD-TOG dataset [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (4 more)

**Figure 4.** Figure 4: The structure of the TD-TOG dataset, which consists of [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 5.** Figure 5: The experimental setup for data collection having a [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 6.** Figure 6: The architecture of Binary-TOG, which is composed of four modules: zero-shot object segmentation (orange), zero-shot object [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

**Figure 7.** Figure 7: Example success and failure cases of Binary-TOG in real-world experiments. For each case, the predictions given by each module [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

6 extracted references · 3 canonical work pages

[216]

One-Shot Transfer of Affordance Regions? AffCorrs!

doi: 10.1177/0278364919872545. Hadjivelichkov, Denis et al. (2022). “One-Shot Transfer of Affordance Regions? AffCorrs!” In:Proceedings of the 2022 Conference on Robot Learning (CoRL). Ed. by Karen Liu, Dana Kulic, and Jeffrey Ichnowski. Vol. 205. Proceedings of Machine Learning Research. PMLR, pp. 550–560. Han, Jiaming et al. (2022). “Align Deep Features...

work page arXiv 2022
[255]

6-DOF GraspNet: Variational Grasp Generation for Object Manipulation

doi: 10.1109/CVPR.2014.39. Mousavian, Arsalan, Clemens Eppner, and Dieter Fox (2019). “6-DOF GraspNet: Variational Grasp Generation for Object Manipulation”. In: 2019 IEEE/CVF International Confer- ence on Computer Vision (ICCV), pp. 2901–2910. doi: 10. 1109/ICCV.2019.00299. Murali, Adithyavairavan et al. (2020). “Same Object, Different Grasps: Data and S...

work page arXiv 2019
[1103]

Learning task-oriented grasping for tool manipulation from simulated self-supervision

doi: 10.1177/02783649231193710. Fang, Kuan et al. (2020). “Learning task-oriented grasping for tool manipulation from simulated self-supervision”. In:The International Journal of Robotics Research39.2-3, pp. 202–

work page doi:10.1177/02783649231193710 2020
[2107]

Reinforcement Learning-Based Grasping via One-Shot Affordance Localization and Zero- Shot Contrastive Language-Image Learning

doi: 10.1007/s00170-018-2363-5. Long, Xiang et al. (2024). “Reinforcement Learning-Based Grasping via One-Shot Affordance Localization and Zero- Shot Contrastive Language-Image Learning”. In: 2024 IEEE/SICE International Symposium on System Integration (SII), pp. 207–212.doi: 10.1109/SII58957.2024.10417178. Margolin, Ran, Lihi Zelnik-Manor, and Ayellet Ta...

arXiv 2024
[3273]

Robust grasping across diverse sensor qualities: The GraspNet-1Billion dataset

doi: 10.1109/IROS.2017.8206162. Fang, Haoshu et al. (2023). “Robust grasping across diverse sensor qualities: The GraspNet-1Billion dataset”. In: The International Journal of Robotics Research42.12, pp. 1094–

arXiv 2023
[9640]

Mechanical Search: Multi- Step Retrieval of a Target Object Occluded by Clutter

doi: 10.1109/ICCV48922.2021.00951. Chinakook (2018). LabelImg2. https://github.com/chinakook/labelImg2. Danielczuk, Michael et al. (2019). “Mechanical Search: Multi- Step Retrieval of a Target Object Occluded by Clutter”. In: 2019 International Conference on Robotics and Automation (ICRA), pp. 1614–1621.doi: 10.1109/ICRA.2019.8794143. Deng, Jia et al. (20...

arXiv 2018

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [216]

One-Shot Transfer of Affordance Regions? AffCorrs!

doi: 10.1177/0278364919872545. Hadjivelichkov, Denis et al. (2022). “One-Shot Transfer of Affordance Regions? AffCorrs!” In:Proceedings of the 2022 Conference on Robot Learning (CoRL). Ed. by Karen Liu, Dana Kulic, and Jeffrey Ichnowski. Vol. 205. Proceedings of Machine Learning Research. PMLR, pp. 550–560. Han, Jiaming et al. (2022). “Align Deep Features...

work page arXiv 2022

[2] [255]

6-DOF GraspNet: Variational Grasp Generation for Object Manipulation

doi: 10.1109/CVPR.2014.39. Mousavian, Arsalan, Clemens Eppner, and Dieter Fox (2019). “6-DOF GraspNet: Variational Grasp Generation for Object Manipulation”. In: 2019 IEEE/CVF International Confer- ence on Computer Vision (ICCV), pp. 2901–2910. doi: 10. 1109/ICCV.2019.00299. Murali, Adithyavairavan et al. (2020). “Same Object, Different Grasps: Data and S...

work page arXiv 2019

[3] [1103]

Learning task-oriented grasping for tool manipulation from simulated self-supervision

doi: 10.1177/02783649231193710. Fang, Kuan et al. (2020). “Learning task-oriented grasping for tool manipulation from simulated self-supervision”. In:The International Journal of Robotics Research39.2-3, pp. 202–

work page doi:10.1177/02783649231193710 2020

[4] [2107]

Reinforcement Learning-Based Grasping via One-Shot Affordance Localization and Zero- Shot Contrastive Language-Image Learning

doi: 10.1007/s00170-018-2363-5. Long, Xiang et al. (2024). “Reinforcement Learning-Based Grasping via One-Shot Affordance Localization and Zero- Shot Contrastive Language-Image Learning”. In: 2024 IEEE/SICE International Symposium on System Integration (SII), pp. 207–212.doi: 10.1109/SII58957.2024.10417178. Margolin, Ran, Lihi Zelnik-Manor, and Ayellet Ta...

arXiv 2024

[5] [3273]

Robust grasping across diverse sensor qualities: The GraspNet-1Billion dataset

doi: 10.1109/IROS.2017.8206162. Fang, Haoshu et al. (2023). “Robust grasping across diverse sensor qualities: The GraspNet-1Billion dataset”. In: The International Journal of Robotics Research42.12, pp. 1094–

arXiv 2023

[6] [9640]

Mechanical Search: Multi- Step Retrieval of a Target Object Occluded by Clutter

doi: 10.1109/ICCV48922.2021.00951. Chinakook (2018). LabelImg2. https://github.com/chinakook/labelImg2. Danielczuk, Michael et al. (2019). “Mechanical Search: Multi- Step Retrieval of a Target Object Occluded by Clutter”. In: 2019 International Conference on Robotics and Automation (ICRA), pp. 1614–1621.doi: 10.1109/ICRA.2019.8794143. Deng, Jia et al. (20...

arXiv 2018