Pith. sign in

REVIEW 3 major objections 6 minor 42 references

Kitchen Robotic Manipulation utilizing Foundation Models

T0 review · 3 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read A modular pipeline of off-the-shelf vision models estimates dishware poses at 89% accuracy without retraining, enabling real sink-to-dishwasher and cup-stacking tasks.

desk verdict Useful modular perception system for kitchen robotics, but the headline ADI is a selected maximum on the same 20 scenes and needs a proper split before it can support the quantitative claims. read the letter →

arxiv 2608.04042 v1 pith:NNNCXATO submitted 2026-08-04 cs.RO

Myung-Hwan Jeon , Sankalp Yamsani , Joohyung Kim This is my paper · ORCID
classification cs.RO
keywords 6Dposeestimationfoundationmodelsopen-vocabularydetectionmulti-viewsegmentation2D-3Dfeaturefusionkitchenroboticsdishwaremanipulationzero-shotperception
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 tries to show that a robot can perceive and handle dishware in real kitchens without being retrained for each new kitchen, by wiring together several pretrained 'foundation' models into one modular pipeline. The pipeline detects objects by text prompt, segments them consistently across camera views, fuses 2D image features with 3D point-cloud features, and registers the results against known CAD models to get 6D poses and grasps. The authors evaluate 24 combinations of swappable models and report that the best one (LLMDet + SAMv2 + DINOv2 + GeoTransformer) reaches an ADI of 89.12% on their 20-scene kitchen dataset, and that a physical robot using it succeeds in over 87% of 296 real manipulation trials across two kitchens. If correct, the result matters because it offers a recipe for building household robots whose perception generalizes by composition of pretrained models rather than by task-specific training.

What carries the argument

The load-bearing mechanism is the modular perception pipeline whose central operation is a training-free 2D–3D feature fusion: for each object instance, patch tokens from a visual foundation model are back-projected onto the point cloud and aggregated by voxel pooling, then L2-normalized and concatenated with point-level features from a geometric foundation model. Classification and coarse pose estimation proceed by cosine-similarity candidate proposal against a database of rendered CAD views, followed by correspondence-based registration in the fused feature space and ICP refinement. The other essential piece is the inventory assumption: because kitchen dishware forms a known, stable set, CAD models and prevalidated grasp configurations constrain the recognition space, which is what lets the pipeline swap foundation models in and out and still produce executable grasps.

What would settle it

Run the best configuration on a new, independently annotated set of kitchen scenes from a third environment (different kitchen, different dishware, same camera and robot) and compute ADI without touching any hyperparameters; if the ADI falls substantially below 89.12%—for instance below the roughly 87% success rate needed for reliable manipulation—the claim of clutter- and occlusion-robust zero-shot generalization would be refuted.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that a system-level integration of open-vocabulary detection, multi-view instance segmentation, instance-aware 3D reconstruction, and 2D–3D feature fusion yields 6D pose estimates for kitchen dishware that are accurate enough for downstream manipulation, with no environment-specific fine-tuning. The best configuration—LLMDet + SAMv2 + DINOv2 + GeoTransformer—achieves 89.12% ADI on the custom 20-scene benchmark under clutter and occlusion, outperforms FoundationPose even when FoundationPose is given ground-truth masks, and transfers to a second kitchen and a second robot arm. The paper interprets the ablation results as showing that image features supply semantic cues, point features supply local geometric structure, and the simple concatenation of L2-normalized features lets each modality compensate for the other's weaknesses.

Load-bearing premise

The load-bearing premise is that the 20-scene custom dataset can serve both as the selection set for the best model and hyperparameters and as the final test set, so the reported 89.12% is a maximum over fitted choices rather than an unbiased estimate of performance on new scenes.

Editorial extensions

If this is right

  • If the central claim holds, household robots can be deployed in new kitchens without per-environment training or fine-tuning, as long as the object inventory and CAD models are known.
  • The modular design means the pipeline can be reconfigured to different compute budgets by substituting lighter or heavier foundation models; the paper shows fusion improves all 24 combinations tested.
  • The 89.12% ADI and over 87% success across 296 trials suggest the approach is strong enough for precision tasks like loading dishwashers and stacking cups, not just simple pick-and-place.
  • The pipeline outperformed FoundationPose on the custom dataset even when FoundationPose was given ground-truth masks, indicating that explicit pose estimation from a fused representation can beat tracking-based approaches on non-sequential inputs.
  • Runtime stays practical for interactive use: roughly 7.7 seconds worst case for a full scene with up to 20 objects on an RTX 4090 workstation.

Reading between the lines

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

  • The same composition recipe—open-vocabulary detection, multi-view segmentation, instance-aware mapping, and 2D–3D feature fusion—could transfer to other inventory-based settings such as warehouses, laboratories, or retail shelves where CAD models of known stock exist; the paper only demonstrates kitchens.
  • Because the best configuration, voxel size, Top-K, and IoU threshold were all selected on the same 20 scenes, the reported 89.12% is an optimistically biased estimate; an independent held-out set of kitchens would be needed to know the true generalization level.
  • The paper's failure analysis attributes most grasp failures to gripper slip and contact with neighboring cups rather than pose error; this suggests that as gripper hardware improves, the perception pipeline may already be near the practical limit for this task class.
  • The strong fusion gains for GeoTransformer—which alone is the weakest geometric model—hint that global-context geometric descriptors are the most complementary to image features; future learned fusion modules could amplify this effect.
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.

Referee Report

3 major / 6 minor

Summary. The paper proposes a modular perception pipeline for kitchen dishware manipulation: open-vocabulary detection (LLMDet or GroundingDINO) with SAMv2 segmentation, multi-view instance association, instance-aware 3D map reconstruction, 2D-3D feature fusion using a visual foundation model (DINOv2, BEiT, or CLIP) and a geometric descriptor (GeDi, FPFH, BufferX, or GeoTransformer), CAD-database object proposal and registration, and grasp planning. The authors evaluate 24 pipeline configurations on a custom 20-scene dataset and report that the best configuration (LLMDet + SAMv2 + DINOv2 + GeoTransformer) achieves an ADI of 89.12%, compare against FoundationPose, and demonstrate real robot execution of sink-to-dishwasher transfer and cup stacking with an aggregate 87.5% success rate over 296 trials. The central claim is that the pipeline generalizes to kitchen scenes without environment-specific retraining.

Significance. If the quantitative results hold, the paper would provide a useful reference integration recipe for zero-shot dishware pose estimation and manipulation, combining systematic component ablation with real-robot validation. Strengths include the breadth of tested model combinations, a clearly described modular architecture, real-world trials across two kitchens, runtime reporting, and the stated intention to release code and supplementary materials. However, the headline accuracy is obtained by selecting both the configuration and the hyperparameters on the same 20-scene dataset, and no held-out evaluation, per-scene scores, confidence intervals, or cross-validation are reported. As a result, the paper's main quantitative claim is not yet credible as an unbiased estimate of performance on new scenes; the qualitative demonstrations and the modular-design contribution remain valuable but do not by themselves support the reported 89.12% figure.

major comments (3)
  1. [§IV-C, Tables I–II, Fig. 6] The headline ADI of 89.12% is the maximum over the 24 configuration scores in Table I (with LLMDet + SAMv2 + DINOv2 + GeoTransformer reaching 88.92% without ICP and 89.12% with ICP), and the same 20-scene dataset is used to choose the voxel size (Fig. 6a), Top-K (Fig. 6b), and IoU threshold (Fig. 6c). Because the configuration and hyperparameters are selected by maximizing ADI or mIoU on the same scenes later used for final scoring, the reported number is a selected maximum, not an unbiased estimate for new kitchen scenes. With n = 20 scenes and no per-scene score distribution or confidence interval, the magnitude of the resulting selection bias cannot be assessed from the manuscript. The authors should recompute the headline result using a held-out validation split or nested cross-validation, report the per-scene ADI distribution, and clearly state which configuration and hyperparameters were fixed before any held-out evaluation.
  2. [§IV-C.1, Table II] The comparison with FoundationPose is not on equal footing: the proposed pipeline runs with its own predicted masks from OVOD + SAMv2, whereas FoundationPose is supplied ground-truth masks, which the text itself describes as an ideal condition unavailable in deployment. The table reports FoundationPose at 87.32 ADI with all ground-truth masks and a failure with the first ground-truth mask, so the 1.80-point gap (89.12 vs. 87.32) may largely reflect the difference in segmentation supervision rather than pose-estimation quality. To support the claim that the best configuration 'achieves higher accuracy than FoundationPose,' either run FoundationPose in its intended tracking mode on suitable sequential data, provide masks from the same upstream detector to both methods, or explicitly frame the result as an upper bound for FoundationPose; in all cases, report uncertainty estimates for both methods.
  3. [§IV-C.1, Table I] Several configuration differences that motivate the 'best-performing configuration' claim are very small, e.g., an 88.86% ADI for DINOv2 + GeDi with fusion versus 88.92% for DINOv2 + GeoTransformer with fusion, and an 88.44% ADI for CLIP + GeoTransformer with fusion. The manuscript reports no per-scene variance, standard errors, or significance tests across the 20 scenes, so these margins are not distinguished from noise. The post-hoc explanation that GeoTransformer's global modeling is 'reinforced' by fusion may be plausible, but the observed differences are too small to establish superiority without paired per-scene results or a significance test.
minor comments (6)
  1. [Algorithm 1] The algorithm title reads '12D-3D Feature Fusion for a Single Instance,' which appears to be a typo for '2D-3D Feature Fusion.'
  2. [Eq. (5)] Equation (5) is typeset incorrectly: the concatenation of the two L2-normalized feature vectors is not legible, and the brackets suggest a row vector that is missing a comma or semicolon separator.
  3. [Eqs. (6)–(8)] The notation Q_m in Eq. (6) is used before being defined; please introduce the query point cloud explicitly in the prose around Eq. (6).
  4. [Table II] The caption of Table II does not explain the column structure or the meanings of 'with All GT Mask' and 'with First GT Mask'; please state operationally what those two conditions are and why the latter is reported as a failure.
  5. [Table I] The table caption says 'without fusion denotes the use of 3D geometric features only,' but the table is organized by visual model rows as well; please clarify which feature streams are active in each block so that readers do not misinterpret the 'without fusion' cases.
  6. [§IV-D.2] In the failure analysis, the sentence about 'heavily cluttered and densely stacked cups' appears to describe the cup-stacking task, but Fig. 7 depicts cup stacking on a counter rather than in the sink; please clarify the environment for that failure mode.

Circularity Check

1 steps flagged · score 6.0 of 10

Headline ADI is selected on the same 20 scenes used for scoring, so the reported 89.12% is a selection-biased maximum rather than an unbiased estimate.

  1. fitted input called prediction [Abstract; Section IV-B; Section IV-C (Tables I-II; Fig. 6)]
    "To validate the pipeline, we built a custom dataset comprising 20 real-world scenes... We evaluated all combinations of two OVOD methods, three visual foundation models, and four geometric foundation models... We therefore adopt 10 mm as the operating voxel size... We therefore set K=3... We therefore set the threshold to 0.6... The best-performing configuration (LLMDet + SAMv2 + DINOv2 + GeoTransformer) achieves an ADI of 89.12% on the 20-scene kitchen benchmark."

    The headline number is not an independent prediction: the same 20-scene dataset is used both to choose the configuration and to choose the hyperparameters, and then to report the chosen configuration's score as an 'achieved' ADI. Table I scores 24 combinations on this dataset, Fig. 6 tunes voxel size, Top-K and IoU threshold on this dataset, and Table II / the abstract then report the selected result. The 89.12% is therefore, by construction, the value that the selection procedure maximized on those scenes, not an out-of-sample estimate; the paper provides no held-out split or per-scene distribution to correct the selection bias. The robot demonstrations and Table III success rates are separate evidence but do not validate the headline ADI.

full rationale

The pipeline itself is a modular composition of externally pretrained foundation models, and there is no derivation chain in which an output is fed back into its own definition. I found no load-bearing self-citation chain: references [5] and [41] are prior hardware and annotation tools by the authors and do not justify the central quantitative claim. The one substantive circularity is statistical: the paper uses the same 20-scene custom dataset both to select the best of 24 combinations and to tune the voxel size, Top-K, and IoU threshold (all chosen by ADI/mIoU on those scenes), and then reports the selected configuration's ADI as the headline result. This makes 89.12% a within-sample selected maximum rather than a predicted performance for new kitchen scenes. The real-robot demonstrations and 296-trial success rates (Table III) are separate, more direct evidence and are not circular; however, they do not rescue the headline ADI from selection bias.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the availability of CAD models and a known object inventory, on the correctness of the registration pipeline, and on the assumption that the custom 20-scene dataset can serve as both selection and test set. There are no invented entities.

free parameters (4)
  • voxel_size = 10 mm
    Chosen on the evaluation dataset via sensitivity analysis (Fig. 6a) to balance ADI and runtime; affects point cloud resolution and feature pooling.
  • top_k = 3
    Chosen on the evaluation dataset via sensitivity analysis (Fig. 6b); controls how many CAD candidates are kept for registration.
  • iou_threshold = 0.6
    Chosen on the evaluation dataset via sensitivity analysis (Fig. 6c); controls cross-view mask association.
  • overlap_epsilon = voxel size
    Distance threshold in overlap ratio equations (Eq. 8) set to the voxel size; tied to the voxel size parameter.
assumptions (3)
  • domain assumption Known CAD inventory and predefined grasps are available for all target objects.
    Section III-B states "we assume that 3D CAD models of the target objects are available and that the object inventory is known", which restricts the pipeline to inventory-based kitchen scenes and is load-bearing for pose and grasp estimation.
  • domain assumption Camera poses from the manipulator can be corrected by coarse registration plus ICP.
    Section III-C2 relies on this correction to build an aligned semantic map; if registration fails, the 3D map misaligns and downstream pose estimates degrade.
  • domain assumption A single text prompt "dishware" is sufficient for the open-vocabulary detector to find all target objects.
    Section III-C1 uses this prompt without per-object or per-scene customization, assuming the detector generalizes across clutter and occlusion.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Kitchen Robotic Manipulation utilizing Foundation Models." pith.science (2026). https://pith.science/paper/NNNCXATO

@misc{pith2026260804042,
  author       = {Pith},
  title        = {Pith review of: Kitchen Robotic Manipulation utilizing Foundation Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NNNCXATO}},
  note         = {Machine review of arXiv:2608.04042}
}
read the original abstract

Deploying robots in everyday human environments requires perception systems that are both robust and adaptable to diverse, dynamic conditions. In this work, we present a modular perception pipeline for household manipulation tasks, with a focus on dishware handling in kitchen environments. The pipeline integrates open-vocabulary object detection, multi-view segmentation, instance-aware 3D reconstruction, and a 2D-3D feature fusion strategy for 6D pose estimation and grasp planning. Its modular design enables systematic substitution of multiple visual and geometric foundation models, allowing us to identify the best-performing configuration through extensive evaluation on a custom kitchen dataset. The best-performing configuration (LLMDet + SAMv2 + DINOv2 + GeoTransformer) achieves an ADI of 89.12\% on the 20-scene kitchen benchmark with cluttered and occluded conditions. Furthermore, real-world demonstrations confirm that the best configuration can be deployed on physical robots without environment-specific retraining, successfully executing tasks such as sink-to-dishwasher transfer and cup stacking. It validates the adaptability and scalability of the pipeline and highlights its potential as a practical framework for household robotic systems. Our code and supplementary materials are available at https://raivlab.github.io/FM_kitchen .

Figures

Figures reproduced from arXiv: 2608.04042 by the authors.

Figure 1
Figure 1. The robot manipulator scans the workspace and [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the perception pipeline. The system integrates Open-Vocabulary Object Detection (OVOD) methods [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Robot system configurations. Three configurations are [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: The qualitative evaluation of the best combination model. The first row illustrates the reconstructed 3D map and [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: The qualitative evaluation of the best model on the [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Performance variation with respect to (a) voxel size, (b) Top-K, and (c) overlap ratio. The highlighted regions indicate the optimal operating ranges, with optima found at voxel size = 10 mm, K = 3, and overlap ratio = 0.6. Task First Kitchen Second Kitchen Cup Stackin…
Figure 7
Figure 7. Figure 7: Tasks performed to demonstrate the best model in a real kitchen environment. [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 27 canonical work pages

  1. [36]

    Freeze: Training-free zero-shot 6d pose estimation with geometric and vision foundation models,

    A. Caraffa, D. Boscaini, A. Hamza, and F. Poiesi, “Freeze: Training-free zero-shot 6d pose estimation with geometric and vision foundation models,” inEuro- pean Conference on Computer Vision. Springer, 2024, pp. 414–431

    work page 2024

  2. [1]

    Tri taking on the hard problems in manipulation research toward making human-assist robots reliable and robust,

    Toyota-Research-Institute, “Tri taking on the hard problems in manipulation research toward making human-assist robots reliable and robust,” 2019. [Online]. Available: https://www.tri.global/news/tr i-taking-hard-problems-manipulation-research-towar d-making-human-assist-robots-reliable/

    work page 2019

  3. [2]

    A robotic system capable of recognition, grasping, and suction for dishwashing automation,

    Y . Fukuzawa, Z. Wang, Y . Mori, and S. Kawamura, “A robotic system capable of recognition, grasping, and suction for dishwashing automation,” in2021 27th In- ternational Conference on Mechatronics and Machine Vision in Practice (M2VIP), 2021, pp. 369–374

    work page 2021

  4. [3]

    Cleaning robots: A review of sensor technologies and intelli- gent control strategies for cleaning,

    R. K. Megalingam, S. R. R. Vadivel, S. S. Kotaprolu, B. Nithul, D. V . Kumar, and G. Rudravaram, “Cleaning robots: A review of sensor technologies and intelli- gent control strategies for cleaning,”Journal of Field Robotics, vol. 42, no. 5, pp. 2234–2259, 2025

    work page 2025

  5. [4]

    Humanoid locomotion and ma- nipulation: Current progress and challenges in control, planning, and learning,

    Z. Gu, J. Li, W. Shen, W. Yu, Z. Xie, S. McCrory, X. Cheng, A. Shamsah, R. Griffin, C. K. Liu, A. Khed- dar, X. B. Peng, Y . Zhu, G. Shi, Q. Nguyen, G. Cheng, H. Gao, and Y . Zhao, “Humanoid locomotion and ma- nipulation: Current progress and challenges in control, planning, and learning,”IEEE/ASME Transactions on Mechatronics, vol. 31, no. 2, pp. 2300–2330, 2026

    work page 2026

  6. [5]

    Papras: Plug-and-play robotic arm system,

    J. Kim, D. C. Mathur, K. Shin, and S. Taylor, “Papras: Plug-and-play robotic arm system,” 2023. [Online]. Available: https://arxiv.org/abs/2302.09655

    arXiv 2023

  7. [6]

    Unitree robotics,

    Unitree-Robotics, “Unitree robotics,” 2024. [Online]. Available: https://www.unitree.com/

    work page 2024

  8. [7]

    Dinov2: Learning robust visual features without supervision,

    M. Oquab, T. Darcet, T. Moutakanni, H. V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. Haziza, F. Massa, A. El-Nouby, M. Assran, N. Ballas, W. Galuba, R. Howes, P.-Y . Huang, S.-W. Li, I. Misra, M. Rabbat, V . Sharma, G. Synnaeve, H. Xu, H. Jegou, J. Mairal, P. Labatut, A. Joulin, and P. Bojanowski, “Dinov2: Learning robust visual features without supe...

    arXiv 2024

Show all 42 references
  1. [8]

    Sam 2: Segment anything in images and videos,

    N. Ravi, V . Gabeur, Y .-T. Hu, R. Hu, C. Ryali, T. Ma, H. Khedr, R. R ¨adle, C. Rolland, L. Gustafson, E. Mintun, J. Pan, K. V . Alwala, N. Carion, C.-Y . Wu, R. Girshick, P. Doll ´ar, and C. Feichtenhofer, “Sam 2: Segment anything in images and videos,” 2024. [Online]. Avail...

    2024 arXiv

  2. [9]

    Dust3r: Geometric 3d vision made easy,

    S. Wang, V . Leroy, Y . Cabon, B. Chidlovskii, and J. Revaud, “Dust3r: Geometric 3d vision made easy,” in 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 20 697–20 709

    2024

  3. [10]

    Grounding dino: Marrying dino with grounded pre- training for open-set object detection,

    S. Liu, Z. Zeng, T. Ren, F. Li, H. Zhang, J. Yang, Q. Jiang, C. Li, J. Yang, H. Su, J. Zhu, and L. Zhang, “Grounding dino: Marrying dino with grounded pre- training for open-set object detection,” 2024. [Online]. Available: https://arxiv.org/abs/2303.05499

    2024 arXiv

  4. [11]

    Beit: Bert pre-training of image transformers,

    H. Bao, L. Dong, S. Piao, and F. Wei, “Beit: Bert pre-training of image transformers,” 2022. [Online]. Available: https://arxiv.org/abs/2106.08254

    2022 arXiv

  5. [12]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever, “Learning transferable visual models from natural language supervision,” 2021. [Online]. Available: https://arxiv.or g/abs/2103.00020

    2021 arXiv

  6. [13]

    Llmdet: Learning strong open-vocabulary object detectors under the supervision of large language models,

    S. Fu, Q. Yang, Q. Mo, J. Yan, X. Wei, J. Meng, X. Xie, and W.-S. Zheng, “Llmdet: Learning strong open-vocabulary object detectors under the supervision of large language models,” 2025. [Online]. Available: https://arxiv.org/abs/2501.18954

    2025 arXiv

  7. [14]

    Learning general and distinctive 3d local deep descriptors for point cloud registration,

    F. Poiesi and D. Boscaini, “Learning general and distinctive 3d local deep descriptors for point cloud registration,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 3, pp. 3979– 3985, 2023

    2023

  8. [15]

    Buffer-x: Towards zero-shot point cloud registration in diverse scenes,

    M. Seo, H. Lim, K. Lee, L. Carlone, and J. Park, “Buffer-x: Towards zero-shot point cloud registration in diverse scenes,” in2025 IEEE/CVF International Conference on Computer Vision (ICCV), 2025, pp. 3851–3862

    2025

  9. [16]

    Geotransformer: Fast and robust point cloud registration with geometric transformer,

    Z. Qin, H. Yu, C. Wang, Y . Guo, Y . Peng, S. Ilic, D. Hu, and K. Xu, “Geotransformer: Fast and robust point cloud registration with geometric transformer,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 8, pp. 9806–9821, 2023

    2023

  10. [17]

    Sam-6d: Segment anything model meets zero-shot 6d object pose esti- mation,

    J. Lin, L. Liu, D. Lu, and K. Jia, “Sam-6d: Segment anything model meets zero-shot 6d object pose esti- mation,” in2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024, pp. 27 906–27 916

    2024

  11. [18]

    Cnos: A strong baseline for cad-based novel object segmentation,

    V . N. Nguyen, T. Groueix, G. Ponimatkin, V . Lepetit, and T. Hodan, “Cnos: A strong baseline for cad-based novel object segmentation,” in2023 IEEE/CVF Inter- national Conference on Computer Vision Workshops (ICCVW), 2023, pp. 2126–2132

    2023

  12. [19]

    Foundationpose: Unified 6d pose estimation and tracking of novel objects,

    B. Wen, W. Yang, J. Kautz, and S. Birchfield, “Foundationpose: Unified 6d pose estimation and tracking of novel objects,” 2024. [Online]. Available: https://arxiv.org/abs/2312.08344

    2024 arXiv

  13. [20]

    Foundationgrasp: Generalizable task-oriented grasping with foundation models,

    C. Tang, D. Huang, W. Dong, R. Xu, and H. Zhang, “Foundationgrasp: Generalizable task-oriented grasping with foundation models,”IEEE Transactions on Au- tomation Science and Engineering, vol. 22, pp. 12 418– 12 435, 2025

    2025

  14. [21]

    Grasp-anything: Large-scale grasp dataset from foundation models,

    A. D. Vuong, M. N. Vu, H. Le, B. Huang, H. T. T. Binh, T. V o, A. Kugi, and A. Nguyen, “Grasp-anything: Large-scale grasp dataset from foundation models,” in 2024 IEEE International Conference on Robotics and Automation (ICRA), 2024, pp. 14 030–14 037

    2024

  15. [22]

    Depth anything: Unleashing the power of large-scale unlabeled data,

    L. Yang, B. Kang, Z. Huang, X. Xu, J. Feng, and H. Zhao, “Depth anything: Unleashing the power of large-scale unlabeled data,” 2024. [Online]. Available: https://arxiv.org/abs/2401.10891

    2024 arXiv

  16. [23]

    Vfmm3d: Releasing the potential of image by vision foundation model for monocular 3d object detection,

    B. Ding, J. Xie, J. Nie, J. Cao, X. Li, and Y . Pang, “Vfmm3d: Releasing the potential of image by vision foundation model for monocular 3d object detection,” 2024. [Online]. Available: https://arxiv.org/abs/2404.09431

    2024 arXiv

  17. [24]

    Mast3r- slam: Real-time dense slam with 3d reconstruction priors,

    R. Murai, E. Dexheimer, and A. J. Davison, “Mast3r- slam: Real-time dense slam with 3d reconstruction priors,” 2025. [Online]. Available: https://arxiv.org/ab s/2412.12392

    2025 arXiv

  18. [25]

    Put it in context with visual foundation models,

    M. McDonald and J. Yu, “Put it in context with visual foundation models,” 2024. [Online]. Available: https://bostondynamics.com/blog/put-it-in-context-wit h-visual-foundation-models/

    2024

  19. [26]

    Innovative integration of visual foundation model with a robotic arm on a mobile platform,

    S. Zhang and Q. Lu, “Innovative integration of visual foundation model with a robotic arm on a mobile platform,” 2024. [Online]. Available: https: //arxiv.org/abs/2404.18720

    2024 arXiv

  20. [27]

    Ok-robot: What really matters in integrating open-knowledge models for robotics,

    P. Liu, Y . Orru, J. Vakil, C. Paxton, N. M. M. Shafiullah, and L. Pinto, “Ok-robot: What really matters in integrating open-knowledge models for robotics,” 2024. [Online]. Available: https://arxiv.org/abs/2401.12202

    2024 arXiv

  21. [28]

    Anygrasp: Robust and efficient grasp perception in spatial and temporal do- mains,

    H.-S. Fang, C. Wang, H. Fang, M. Gou, J. Liu, H. Yan, W. Liu, Y . Xie, and C. Lu, “Anygrasp: Robust and efficient grasp perception in spatial and temporal do- mains,”IEEE Transactions on Robotics, vol. 39, no. 5, pp. 3929–3945, 2023

    2023

  22. [29]

    Tidybot: Personalized robot assistance with large lan- guage models,

    J. Wu, R. Antonova, A. Kan, M. Lepert, A. Zeng, S. Song, J. Bohg, S. Rusinkiewicz, and T. Funkhouser, “Tidybot: Personalized robot assistance with large lan- guage models,”Autonomous Robots, vol. 47, no. 8, pp. 1087–1102, 2023

    2023

  23. [30]

    V oxposer: Composable 3d value maps for robotic manipulation with language models,

    W. Huang, C. Wang, R. Zhang, Y . Li, J. Wu, and L. Fei-Fei, “V oxposer: Composable 3d value maps for robotic manipulation with language models,” 2023. [Online]. Available: https://arxiv.org/abs/2307.05973

    2023 arXiv

  24. [31]

    Rt-2: Vision-language-action models transfer web knowledge to robotic control,

    B. Zitkovich, T. Yu, S. Xu, P. Xu, T. Xiao, F. Xia, J. Wu, P. Wohlhart, S. Welker, A. Wahid, Q. Vuong, V . Vanhoucke, H. Tran, R. Soricut, A. Singh, J. Singh, P. Sermanet, P. R. Sanketi, G. Salazar, M. S. Ryoo, K. Reymann, K. Rao, K. Pertsch, I. Mordatch, H. Michalewski, Y . L...

    2023

  25. [32]

    Openvla: An open-source vision-language- action model,

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. Foster, G. Lam, P. Sanketi, Q. Vuong, T. Kollar, B. Burchfiel, R. Tedrake, D. Sadigh, S. Levine, P. Liang, and C. Finn, “Openvla: An open-source vision-language- action model,” 2024. [Onlin...

    2024 arXiv

  26. [33]

    Open x-embodiment: Robotic learning datasets and rt-x models : Open x-embodiment collaboration0,

    A. O’Neill, A. Rehman, A. Maddukuri, A. Gupta, A. Padalkar, A. Lee, A. Pooley, A. Gupta, A. Man- dlekar, A. Jain, A. Tung, A. Bewley, A. Herzog, A. Irpan, A. Khazatsky, A. Rai, A. Gupta, A. Wang, A. Singh, A. Garg, A. Kembhavi, A. Xie, A. Brohan, A. Raffin, A. Sharma, A. Yavar...

    2024

  27. [34]

    Octo: An open-source generalist robot policy,

    O. M. Team, D. Ghosh, H. Walke, K. Pertsch, K. Black, O. Mees, S. Dasari, J. Hejna, T. Kreiman, C. Xu, J. Luo, Y . L. Tan, L. Y . Chen, P. Sanketi, Q. Vuong, T. Xiao, D. Sadigh, C. Finn, and S. Levine, “Octo: An open-source generalist robot policy,” 2024. [Online]. Available: ...

    2024 arXiv

  28. [35]

    Kiss-matcher: Fast and ro- bust point cloud registration revisited,

    H. Lim, D. Kim, G. Shin, J. Shi, I. Vizzo, H. Myung, J. Park, and L. Carlone, “Kiss-matcher: Fast and ro- bust point cloud registration revisited,” in2025 IEEE International Conference on Robotics and Automation (ICRA), 2025, pp. 11 104–11 111

    2025

  29. [37]

    Robot sensor calibration: solv- ing ax=xb on the euclidean group,

    F. Park and B. Martin, “Robot sensor calibration: solv- ing ax=xb on the euclidean group,”IEEE Transactions on Robotics and Automation, vol. 10, no. 5, pp. 717– 721, 1994

    1994

  30. [38]

    The pinoc- chio c++ library : A fast and flexible implementation of rigid body dynamics algorithms and their analytical derivatives,

    J. Carpentier, G. Saurel, G. Buondonno, J. Mirabel, F. Lamiraux, O. Stasse, and N. Mansard, “The pinoc- chio c++ library : A fast and flexible implementation of rigid body dynamics algorithms and their analytical derivatives,” in2019 IEEE/SICE International Sympo- sium on Syst...

    2019

  31. [39]

    Rrt-connect: An efficient approach to single-query path planning,

    J. Kuffner and S. LaValle, “Rrt-connect: An efficient approach to single-query path planning,” inProceed- ings 2000 ICRA. Millennium Conference. IEEE In- ternational Conference on Robotics and Automation. Symposia Proceedings (Cat. No.00CH37065), vol. 2, 2000, pp. 995–1001 vol.2

    2000

  32. [40]

    Fcl: A general purpose library for collision and proximity queries,

    J. Pan, S. Chitta, and D. Manocha, “Fcl: A general purpose library for collision and proximity queries,” in 2012 IEEE International Conference on Robotics and Automation, 2012, pp. 3859–3866

    2012

  33. [41]

    Transpose: Large-scale multispectral dataset for transparent object,

    J. Kim, M.-H. Jeon, S. Jung, W. Yang, M. Jung, J. Shin, and A. Kim, “Transpose: Large-scale multispectral dataset for transparent object,”The International Jour- nal of Robotics Research, vol. 43, no. 6, pp. 731–738, 2024

    2024

  34. [42]

    Fin ray® effect inspired soft robotic gripper: From the robosoft grand challenge toward optimization,

    W. Crooks, G. Vukasin, M. O’Sullivan, W. Messner, and C. Rogers, “Fin ray® effect inspired soft robotic gripper: From the robosoft grand challenge toward optimization,”Frontiers in Robotics and AI, vol. 3, p. 70, 2016

    2016

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.