Pith. sign in

REVIEW 2 major objections 5 minor 1 cited by

Amodal 3D Reconstruction for Robotic Manipulation via Stability and Connectivity

T0 review · 2 major / 5 minor · reviewed 2026-08-27 · deepseek-v4-flash

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

Pith's one-line read Adding stability and connectivity priors to 3D reconstruction boosts robot manipulation success by 42%.

desk verdict A promising empirical recipe for physics-aware 3D reconstruction whose central gradient derivation has a real gap; the results are interesting but the stated mechanism is not yet proven. read the letter →

arxiv 2009.13146 v1 pith:METRVJIT submitted 2020-09-28 cs.RO cs.CVcs.LG

William Agnew , Christopher Xie , Aaron Walsman , Octavian Murad , Caelen Wang , Pedro Domingos , Siddhartha Srinivasa This is my paper · ORCID
classification cs.ROcs.CVcs.LG
keywords 3Dreconstructionamodalroboticmanipulationstabilitypriorconnectivitymodel-basedcontrolclutteredscenesvoxelrepresentation
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 argues that existing 3D object reconstruction networks, which optimize visual fidelity via chamfer distance or voxel IOU, produce meshes with poor physical realism in cluttered scenes, and that these physically implausible meshes cause model-based manipulation to fail. It proposes ARM, which adds a differentiable stability prior (reconstructions should be in static equilibrium), a connectivity prior (each object should be one connected component), and a multi-channel voxel input that lets the network see other objects' occupancy and unobserved space. On a newly introduced cluttered benchmark, ARM reduces chamfer loss by 28% and raises manipulation success on unseen objects by 42% over the GenRE baseline. The central message is that physical priors, not visual metrics, are what make reconstructions useful for control.

What carries the argument

The central machinery is a four-channel voxel representation plus two differentiable physics-based losses. The input encodes, per object, its own occupancy, the occupancy of all other objects, empty space, and unobserved space; the losses are a stability loss, defined as the probability that the sampled voxel grid is in static equilibrium (center of mass within the base of support along discretized horizontal directions), and a connectivity loss, defined as the probability that every occupied voxel pair is connected by at least one path. Both losses are made tractable by approximating the probability of stability or connectivity and taking derivatives with respect to voxel occupancy probabilities, so the whole pipeline trains end-to-end.

What would settle it

Reconstruct a scene containing an object that is observably not in static equilibrium, such as a bottle leaning on a second object or an object captured mid-fall, and measure chamfer distance to the true mesh. If ARM's reconstruction is further from the true shape than a baseline without the stability prior on such scenes, the prior is actively distorting rather than correcting shape estimation; the paper's own evaluation excludes this case because synthetic objects are always settled before rendering.

Watch

Extended reading notes

Core claim

The paper's central claim is that optimizing reconstruction for physical stability and connectivity, rather than purely for visual similarity, yields 3D models that are both quantitatively better (28% lower chamfer distance) and substantially better for downstream manipulation, raising success rates on grasping, pushing, and rearrangement tasks with unseen objects by 42% over GenRE. The reconstruction network is trained with two auxiliary differentiable losses: a stability loss that penalizes any voxel configuration whose center of mass falls outside its base of support, and a connectivity loss that rewards configurations in which every occupied voxel pair is joined by a high-probability path. Applied only to occluded voxels, these losses guide the network to infer hidden bases and fill gaps between visible fragments. The paper further shows that the gains concentrate at high occlusion: at 10% target visibility, ARM retains 75% of its unoccluded success rate versus 25% for the baseline.

Load-bearing premise

The load-bearing premise is that the scene and all objects in it are at rest under gravity before manipulation, so that the stability prior's target — static equilibrium — matches the ground truth.

Editorial extensions

If this is right

  • Reconstructions of heavily occluded objects become usable for model-based control: success at 10% visibility stays at 75% of the unoccluded rate, versus 25% without the priors.
  • The stability prior alone brings reconstructed scenes' physics close to ground truth meshes in simulation, while visual-only baselines frequently tumble.
  • The framework is modular: any amodal reconstruction network can accept the four-channel input and the two losses.
  • Physical priors improve not just chamfer distance but task success across grasping, pushing, and rearrangement, suggesting that control-relevant fidelity is learnable.

Reading between the lines

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

  • If the stability prior is applied to scenes where an object is genuinely not at rest (for example, mid-manipulation or leaning), it would likely bias the reconstruction toward a stable shape that misrepresents true geometry; a direct test is to run ARM on non-equilibrium scenes and compare shape error.
  • The same template — a differentiable prior over a physical property, applied only to unobserved voxels — could extend to friction, mass distribution, compliance, or kinematic articulation, each of which would inject a different control-relevant constraint.
  • The multi-channel input suggests that explicitly communicating inter-object occupancy during reconstruction is a cheap way to improve amodal completion; future work might swap the hand-crafted channels for learned relational features.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Request a human review

A listed scientist reviews the paper for a fee and the review publishes here regardless of verdict. See the reviewers or get listed.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 5 minor

Summary. This paper presents ARM, an amodal 3D reconstruction system for cluttered tabletop scenes, built on the GenRE reconstruction network. Its contributions are two differentiable auxiliary losses: a stability prior that penalizes reconstructions whose center of mass is not supported, and a connectivity prior that penalizes disconnected predicted objects, together with a four-channel voxel input representation that encodes the presence of other objects and visibility. The system is trained on synthetic MuJoCo scenes with ShapeNet objects and evaluated on held-out objects both by reconstruction metrics (chamfer distance, simulated scene stability) and by three simulated manipulation tasks (grasping, pushing, rearrangement). The paper reports a 28% improvement in reconstruction quality and a 42% improvement in manipulation success over the GenRE baseline.

Significance. The empirical study is substantial: about 2.3 million training instances, a separate held-out reconstruction benchmark, a 2,574-task manipulation suite, targeted ablations, and a promised code release. If the physical priors work as described, the central idea is valuable and timely: reconstruction objectives such as chamfer distance are indeed poorly aligned with later physical interaction, and differentiable stability/connectivity regularizers are a modular way to address this. The paper also provides a clear decomposition of the gains across occlusion levels and gives qualitative evidence that the inferred bases and connections are physically meaningful. However, the mechanistic attribution of the empirical gains to the stated objectives currently rests on gradient derivations in the appendices that contain mathematical errors. The empirical results may still be valid as an engineering contribution, but the paper does not yet establish that the reported improvements are caused by the stated stability and connectivity losses.

major comments (2)
  1. [Appendix A, Eqs. (3)-(5)] The stability-gradient derivation is not valid as written. The appendix asserts that d log P(i_s > M_s(v)) / d V_o(i) = 0, with the reasoning that adding (or increasing the chance of) a number in a set cannot change whether the mean of that set is greater than, equal to, or less than that number. This is false: M_s(v) is the center of mass of the random binary voxel configuration v, and its distribution depends on V_o. Increasing V_o(i) changes both the numerator and the denominator of the center-of-mass ratio, so the CDF term P(i_s > M_s(v)) has a nonzero derivative with respect to V_o(i). That omitted derivative multiplies the same product that defines u'_s and can change the sign or magnitude of the stability gradient. The same appendix also calls P(i_s > M_s(v)) the CDF of a linear combination of Bernoulli variables, but the center of mass is, for equal-mass voxels, the ratio (sum_i i_s v_i)/(sum_i v_i), not a linear combination; the normal approximation to the Poisson-binomial CDF therefore does not apply to M_s(v) as stated. Since the stability prior is the paper's first and most heavily credited contribution, these issues prevent attributing the reported 28% chamfer and 42% manipulation-success gains to the stated objective. The authors should derive the correct gradient (including the CDF derivative), justify or replace the ratio approximation, and ideally retrain with the corrected loss to confirm the empirical improvements.
  2. [Sec. 3.2.2, Eq. (7)] The connectivity gradient is also incomplete. Eq. (7) sums only over triples with a, b, and c all distinct, but the derivative of the product in Eq. (6) with respect to V_o(c) also contains terms for pairs (c, b) and (a, c), where the occupancy factor V_o(a)V_o(b) itself depends on V_o(c). These endpoint terms are not negligible in general: the probability of any path from c to b includes V_o(c) as a factor, so dP(C(c,b))/dV_o(c) is nonzero. As written, Eq. (7) is not the derivative of Eq. (6), even under the most-likely-path approximation. The authors should either correct the endpoint terms or explicitly state and justify the additional approximation that drops them; a retraining experiment with the corrected connectivity gradient would clarify whether this affects the reported manipulation improvements.
minor comments (5)
  1. [Sec. 3.2.1 / Sec. 4.1] The stability prior assumes that the scene and its objects are at rest before manipulation. This premise is satisfied by construction in the synthetic benchmark, where objects are dropped and allowed to settle before rendering, but the paper does not discuss what happens when the premise is violated, such as a leaning object or an object being manipulated at the moment of observation. A brief limitation paragraph and, if feasible, a sensitivity experiment with non-resting objects would strengthen the paper.
  2. [Sec. 4.2 / Fig. 4] The multi-channel input representation is listed as a contribution, but the text does not numerically separate its effect from the two physical losses. Since ARM-C-S is trained without either loss but with the multi-channel input, reporting the explicit ARM-C-S-versus-baseline chamfer and success numbers would let the reader quantify how much of the 28%/42% gains come from the input representation alone.
  3. [Sec. 3.1 / Sec. 4.4] The phrase 'execute this plan in the real environment' is ambiguous: the manipulation evaluation appears to execute trajectories in the ground-truth MuJoCo environment, not on a physical robot. Please clarify that all reported manipulation results are simulated and that no hardware experiments were performed.
  4. [Appendix A] The sentence 'the cdf of which is inefficient to compute exactly [cite]' contains a dangling placeholder '[cite]' instead of an actual citation. Please add the appropriate reference for the Poisson-binomial CDF.
  5. [Appendix B / Sec. 3.2.2] The notation for the most-likely path is inconsistent between the main text (t*) and the appendix (ts), and the appendix uses both h'_s and h_s for the same quantity in the stability derivation. Please unify the notation.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: stability and connectivity priors are training regularizers, and the reported reconstruction and manipulation gains are measured on held-out objects with independent metrics.

full rationale

ARM's central claims are empirical: adding differentiable stability and connectivity losses to a GenRE-style reconstruction network improves Chamfer distance on held-out objects and manipulation success on unseen YCB and internet-repository objects. The stability and connectivity losses are training regularizers over predicted voxel probabilities, not fitted predictors of the reported evaluation metrics. The reported metrics—Chamfer distance, MuJoCo-simulated manipulation success, and L2 displacement under gravity—are computed from reconstructed meshes in a physics engine, not from the loss equations themselves. The Figure 4 stability metric is related to the stability prior, but the loss (Eq. 1-2) is a probabilistic static-equilibrium condition, whereas the evaluation is dynamical simulation displacement; the latter is an independent proxy, not Eq. 1 by construction. No load-bearing claim rests on a self-citation: GenRE and UOIS-Net are external baselines/components, and no uniqueness theorem or ansatz is imported from the authors' prior work. Appendix A's assertion that d log P(i_s > M_s(v))/dV_o(i)=0 and the '[cite]' placeholder are correctness/completeness concerns, not circularity, because they concern whether the implemented gradient matches the stated objective, not whether the objective is defined in terms of its own output. Thus no circular step is exhibited.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central contribution does not rest on newly fitted physical constants. The hand-chosen discretization and threshold parameters (|S|=25, 0.5 binarization, k=4, coarsening 8) shape the loss landscape, and the physics priors rely on the domain assumptions of static stability, single-component connectivity, and simulator fidelity. The most fragile axiom is the claimed zero derivative of the center-of-mass CDF, used to simplify the stability gradient. No invented entities are introduced.

free parameters (4)
  • direction count |S| = 25
    The stability loss discretizes the continuous set of directions perpendicular to gravity into 25 evenly spaced directions. This is a hand-chosen approximation that the gradient depends on.
  • binarization threshold = 0.5
    The stability and connectivity losses assume a voxel exists if its predicted probability is at least 0.5. This threshold is used in the derivative approximations (Appendix A and B).
  • voxel grid side factor k = 4
    The per-object voxel grid is centered at the object's center of mass with side length k*delta_o, with k=4, to allow filling occluded regions. Hand-chosen.
  • connectivity coarsening factor = 8
    The predicted voxel grid is coarsened by a factor of 8 during connectivity loss computation for efficiency, affecting which paths are considered.
assumptions (6)
  • domain assumption Objects in the scene are stable before manipulation
    Implicit in the stability prior in Section 3.2.1; the loss penalizes reconstructions that are not in static equilibrium. If the scene contains objects not at rest, the prior is inappropriate.
  • domain assumption Each object is a single connected component
    The connectivity loss (Section 3.2.2) assumes each reconstructed object should be connected; this may not hold for genuinely multi-part objects or mis-segmentations.
  • domain assumption MuJoCo's contact simulation is an accurate model of real physics for the objects and tasks
    All experiments use MuJoCo as the ground-truth environment for stability measurement and task success. If MuJoCo's physics is unrealistic, the reported manipulation improvements may not transfer.
  • ad hoc to paper The normal approximation to the Poisson Binomial CDF is sufficiently accurate
    Used in Appendix A to make P(is > Ms(v)) tractable; introduces approximation error in the stability gradient.
  • ad hoc to paper Most-likely-path approximation for connectivity is sufficient
    Appendix B replaces the union of all paths between two voxels with the single most likely path and the most likely path through the queried voxel; this is a heuristic assumption.
  • ad hoc to paper d/dV_o(i) P(i_s > M_s(v)) = 0
    Assumed in Appendix A to simplify the stability gradient. The reasoning given is not generally correct: adding a voxel can shift the center of mass and change this probability.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Amodal 3D Reconstruction for Robotic Manipulation via Stability and Connectivity." pith.science (2026). https://pith.science/paper/METRVJIT

@misc{pith2026200913146,
  author       = {Pith},
  title        = {Pith review of: Amodal 3D Reconstruction for Robotic Manipulation via Stability and Connectivity},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/METRVJIT}},
  note         = {Machine review of arXiv:2009.13146}
}
read the original abstract

Learning-based 3D object reconstruction enables single- or few-shot estimation of 3D object models. For robotics, this holds the potential to allow model-based methods to rapidly adapt to novel objects and scenes. Existing 3D reconstruction techniques optimize for visual reconstruction fidelity, typically measured by chamfer distance or voxel IOU. We find that when applied to realistic, cluttered robotics environments, these systems produce reconstructions with low physical realism, resulting in poor task performance when used for model-based control. We propose ARM, an amodal 3D reconstruction system that introduces (1) a stability prior over object shapes, (2) a connectivity prior, and (3) a multi-channel input representation that allows for reasoning over relationships between groups of objects. By using these priors over the physical properties of objects, our system improves reconstruction quality not just by standard visual metrics, but also performance of model-based control on a variety of robotics manipulation tasks in challenging, cluttered environments. Code is available at github.com/wagnew3/ARM.

Figures

Figures reproduced from arXiv: 2009.13146 by the authors.

Figure 1
Figure 1. Comparison of physical behaviors of reconstructions from different algorithms. The base [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. A visual representation of our modular framework. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Impact of stability and connectivity objectives. Left: occupancy probabilities of an esti [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: (left) Chamfer distances on held-out objects, broken down by observation occlusion. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Qualitative reconstruction results. ARM is able to infer both bases and occluded object [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: (left) Success rates on manipulation tasks using models generated by different reconstruc [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Exploiting Radiance Fields for Grasp Generation on Novel Synthetic Views

    cs.RO 2025-05 conditional novelty 6.0 of 10

    Rendering 16 novel views from a radiance field built from 3 real views adds force-closure grasps and improves grasp coverage on GraspNet scenes.

Reference graph

Works this paper leans on

41 extracted references · 40 canonical work pages · cited by 1 Pith paper

  1. [1]

    Computer poker: A review

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017

    work page Pith review arXiv 2017

  2. [2]

    Levine, P

    S. Levine, P. Pastor, A. Krizhevsky, J. Ibarz, and D. Quillen. Learning hand-eye coordination for robotic grasping with deep learning and large-scale data collection. The International Journal of Robotics Research, 37(4-5):421–436, 2018

    work page 2018

  3. [3]

    Calli, A

    B. Calli, A. Singh, A. Walsman, S. Srinivasa, P. Abbeel, and A. M. Dollar. The ycb object and model set: Towards common benchmarks for manipulation research. In International Conference on Advanced Robotics (ICAR) , 2015

    work page 2015

  4. [4]

    Xiang, T

    Y . Xiang, T. Schmidt, V . Narayanan, and D. Fox. Posecnn: A convolutional neural network for 6d object pose estimation in cluttered scenes. Robotics: Science and Systems (RSS) , 2018

    work page 2018

  5. [5]

    Y . Li, G. Wang, X. Ji, Y . Xiang, and D. Fox. Deepim: Deep iterative matching for 6d pose estimation. In European Conference on Computer Vision (ECCV), 2018

    work page 2018

  6. [6]

    Tremblay, T

    J. Tremblay, T. To, B. Sundaralingam, Y . Xiang, D. Fox, and S. Birchfield. Deep object pose estimation for semantic robotic grasping of household objects. In Conference on Robot Learn- ing (CoRL), 2018

    work page 2018

  7. [7]

    Coumans and Y

    E. Coumans and Y . Bai. Pybullet, a python module for physics simulation for games, robotics and machine learning. GitHub repository, 2016

    work page 2016

  8. [8]

    Todorov, T

    E. Todorov, T. Erez, and Y . Tassa. Mujoco: A physics engine for model-based control. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , 2012

    work page 2012

Show all 41 references
  1. [9]

    Dogar and S

    M. Dogar and S. Srinivasa. A framework for push-grasping in clutter. Robotics: Science and Systems (RSS), 2011

    2011

  2. [10]

    Stilman, J.-U

    M. Stilman, J.-U. Schamburek, J. Kuffner, and T. Asfour. Manipulation planning among mov- able obstacles. In IEEE International Conference on Robotics and Automation (ICRA) , 2007

    2007

  3. [11]

    Williams, A

    G. Williams, A. Aldrich, and E. Theodorou. Model predictive path integral control using covariance variable importance sampling. arXiv preprint arXiv:1509.01149, 2015

    2015 arXiv

  4. [12]

    R. A. Newcombe, S. Izadi, O. Hilliges, D. Molyneaux, D. Kim, A. J. Davison, P. Kohi, J. Shot- ton, S. Hodges, and A. Fitzgibbon. Kinectfusion: Real-time dense surface mapping and track- ing. In IEEE International Symposium on Mixed and Augmented Reality (ISMAR) , 2011

    2011

  5. [13]

    Gkioxari, J

    G. Gkioxari, J. Malik, and J. Johnson. Mesh r-cnn. In IEEE International Conference on Computer Vision (ICCV), 2019

    2019

  6. [14]

    Zhang, Z

    X. Zhang, Z. Zhang, C. Zhang, J. B. Tenenbaum, W. T. Freeman, and J. Wu. Learning to Reconstruct Shapes from Unseen Classes. In Advances in Neural Information Processing Systems (NeurIPS), 2018

    2018

  7. [15]

    Smith, S

    E. Smith, S. Fujimoto, and D. Meger. Multi-view silhouette and depth decomposition for high resolution 3d object representation. In Advances in Neural Information Processing Systems (NeurIPS), 2018. 9

    2018

  8. [16]

    Mescheder, M

    L. Mescheder, M. Oechsle, M. Niemeyer, S. Nowozin, and A. Geiger. Occupancy networks: Learning 3d reconstruction in function space. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019

    2019

  9. [17]

    Tulsiani, S

    S. Tulsiani, S. Gupta, D. F. Fouhey, A. A. Efros, and J. Malik. Factoring shape, pose, and layout from the 2d image of a 3d scene. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018

    2018

  10. [18]

    Y . Nie, X. Han, S. Guo, Y . Zheng, J. Chang, and J. J. Zhang. Total3dunderstanding: Joint layout, object pose and mesh reconstruction for indoor scenes from a single image. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2020

    2020

  11. [19]

    Kulkarni, I

    N. Kulkarni, I. Misra, S. Tulsiani, and A. Gupta. 3d-relnet: Joint object and relational network for 3d prediction. In IEEE International Conference on Computer Vision (ICCV) , 2019

    2019

  12. [20]

    X. Yan, J. Yang, E. Yumer, Y . Guo, and H. Lee. Perspective transformer nets: Learning single- view 3d object reconstruction without 3d supervision. In Advances in Neural Information Processing Systems (NeurIPS), 2016

    2016

  13. [21]

    D. Shin, C. C. Fowlkes, and D. Hoiem. Pixels, voxels, and views: A study of shape represen- tations for single view 3d object shape prediction. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018

    2018

  14. [22]

    X. Han, H. Laga, and M. Bennamoun. Image-based 3d object reconstruction: State-of-the- art and trends in the deep learning era. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2019

    2019

  15. [23]

    Kundu, Y

    A. Kundu, Y . Li, and J. M. Rehg. 3d-rcnn: Instance-level 3d object reconstruction via render- and-compare. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2018

    2018

  16. [24]

    Sucar, K

    E. Sucar, K. Wada, and A. Davison. Neural object descriptors for multi-view shape reconstruc- tion. arXiv preprint arXiv:2004.04485, 2020

    2004 arXiv

  17. [25]

    Z. Jia, A. Gallagher, A. Saxena, and T. Chen. 3d-based reasoning with blocks, support, and stability. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , 2013

    2013

  18. [26]

    Y . Du, Z. Liu, H. Basevi, A. Leonardis, B. Freeman, J. Tenenbaum, and J. Wu. Learning to exploit stability for 3d scene parsing. In Advances in Neural Information Processing Systems (NeurIPS), 2018

    2018

  19. [27]

    Y . Chen, S. Huang, T. Yuan, S. Qi, Y . Zhu, and S.-C. Zhu. Holistic++ scene understanding: Single-view 3d holistic scene parsing and human pose estimation with human-object inter- action and physical commonsense. In IEEE International Conference on Computer Vision (ICCV), 2019

    2019

  20. [28]

    Varley, C

    J. Varley, C. DeChant, A. Richardson, J. Ruales, and P. Allen. Shape completion enabled robotic grasping. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2017

    2017

  21. [29]

    Van der Merwe, Q

    M. Van der Merwe, Q. Lu, B. Sundaralingam, M. Matak, and T. Hermans. Learning continuous 3d reconstructions for geometrically aware grasping. In IEEE Conference on Robotics and Automation (ICRA), 2020

    2020

  22. [30]

    X. Yan, J. Hsu, M. Khansari, Y . Bai, A. Pathak, A. Gupta, J. Davidson, and H. Lee. Learning 6-dof grasping interaction via deep geometry-aware 3d representations. In IEEE Conference on Robotics and Automation (ICRA) , 2018

    2018

  23. [31]

    Lundell, F

    J. Lundell, F. Verdoja, and V . Kyrki. Robust grasp planning over uncertain shape completions. In IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , 2019

    2019

  24. [32]

    Lundell, F

    J. Lundell, F. Verdoja, and V . Kyrki. Beyond top-grasps through scene completion. In IEEE Conference on Robotics and Automation (ICRA) , 2020. 10

    2020

  25. [33]

    C. Xie, Y . Xiang, A. Mousavian, and D. Fox. The best of both modes: Separately leveraging rgb and depth for unseen object instance segmentation. In Conference on Robot Learning (CoRL), 2019

    2019

  26. [34]

    W. E. Lorensen and H. E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. ACM siggraph computer graphics, 21(4):163–169, 1987

    1987

  27. [35]

    Mamou, E

    K. Mamou, E. Lengyel, and A. Peters. V olumetric hierarchical approximate convex decompo- sition. Game Engine Gems 3 , pages 141–158, 2016

    2016

  28. [36]

    P. P. Urone, K. Dirks, and M. Sharma. Statics and Torque, page 289316. OpenStax

  29. [37]

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

    2015 arXiv

  30. [38]

    Mahler, J

    J. Mahler, J. Liang, S. Niyaz, M. Laskey, R. Doan, X. Liu, J. A. Ojea, and K. Goldberg. Dex- net 2.0: Deep learning to plan robust grasps with synthetic point clouds and analytic grasp metrics. arXiv preprint arXiv:1703.09312, 2017

    2017 arXiv

  31. [39]

    Lowrey, A

    K. Lowrey, A. Rajeswaran, S. Kakade, E. Todorov, and I. Mordatch. Plan Online, Learn Offline: Efficient Learning and Exploration via Model-Based Control. In ICLR, 2019. A Stability Loss Derivation Recall that Vo parameterizes a multivariate Bernoulli distribution over binarized ...

    2019

  32. [40]

    Selected shapenet training categories: bag, traveling bag, travelling bag, grip, suitcase, bird- house, bottle, bowl, camera, photographic camera, can, tin, tin can, cap, clock, computer keyboard, keypad, dishwasher, dish washer, dishwashing machine, helmet, jar, knife, laptop...

  33. [41]

    The reward function had three states: ungrasped, grasping, and grasped

    Target manipulation objects: YCB Objects: master chef can, cracker box, sugar box, tomato soup can, mustard bottle, apple, orange, pitcher base, bleach cleanser, bowl, mug, wood block, tennis ball, rubiks cube Objects from internet repository: cup (x2), glass (x3), vase, lamp ...

Pith tools

Reviewed August 27, 2026 · model on record in the stance chip above.