Pith. sign in

REVIEW 5 major objections 6 minor 35 references

Manipulating Elasto-Plastic Objects With 3D Occupancy and Learning-Based Predictive Control

T0 review · 5 major / 6 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 This paper claims that dense 3D occupancy, learned dynamics, and predictive control let a robot shape plasticine into given goal shapes.

desk verdict Dense 3D occupancy plus learned MPC is a solid step for elasto-plastic manipulation, but the 'given goal shape' claim outruns what the representation and experiments support. read the letter →

arxiv 2505.16249 v2 pith:FSBUMVPD submitted 2025-05-22 cs.RO cs.AI

Zhen Zhang , Xiangyu Chu , Yunxi Tang , Lulu Zhao , Jing Huang , Zhongliang Jiang , K. W. Samuel Au This is my paper
classification cs.ROcs.AI
keywords elasto-plasticobjectmanipulation3Doccupancyrepresentationdeformabledynamicsmodelpredictivecontrolgraphneuralnetworkconvolutionalplasticineshapingmulti-viewRGBprediction
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

This paper tries to establish that a robot can shape an elasto-plastic object such as plasticine into a specified 3D goal shape by representing the object as a dense voxel grid of 3D occupancy rather than as a sparse point cloud or surface mesh. The authors claim that with a neural network that predicts 3D occupancy from multi-view RGB images, a learned dynamics model built on a 3D CNN and a graph neural network, and a model-predictive controller initialized from shape geometry, a two-finger gripper can deform the material into letters and layered shapes in both simulation and the real world. The point of the work is that 3D occupancy is dense, volumetric, and semantic enough to support both accurate dynamics prediction and goal-conditioned planning for an object class that is notoriously hard to represent. If the claim holds, it offers a general recipe for deformable-object manipulation that does not require accurate real-time mesh reconstruction or particle simulation.

What carries the argument

The central object is the dense 3D occupancy grid, a voxelized representation in which each voxel holds the probability that it is occupied by the plasticine, the gripper, or the operating plane. From this grid the pipeline keeps a sparse voxel state via farthest point sampling, forms a graph whose edges encode spatial and finger-object relations, and feeds the graph into a dynamics model that combines a 3D sparse CNN for multi-scale volumetric features with a graph neural network for node-state propagation. A shape-based action initialization module uses the geometric mismatch between the current and goal states to seed the MPC action sequence, and L-BFGS optimizes the actions against the learned dynamics model. The occupancy-prediction network, supervised by the offline-generated occupancy labels, supplies the inferred state that lets the whole loop run at test time from RGB images alone.

What would settle it

Run the framework on a task that forces a topological change, for example splitting one clay blob into two separate blobs or merging two blobs into one, and compare the reconstructed occupancy ground truth and the final shaped object against the goal; if the occupancy labels miss the split or merge, the learned dynamics and controller cannot plan it, and the general 'given goal shape' claim fails.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the combination of a 3D-occupancy state representation, a 3D CNN-GNN dynamics model, and learning-based model predictive control successfully shapes elasto-plastic objects into given goal shapes. The authors report that their 3D-occupancy-based state representation beats patch-based and crop-based alternatives on Earth Mover's Distance, Density-aware Chamfer Distance, and Chamfer Distance, and that their dynamics model beats two GNN-only baselines on the same metrics. They further demonstrate in hardware that the framework can produce single-layer letters with varying symmetry, two-layer stamps with controlled height ratios, and shapes starting from different initial geometries, colors, and materials. The paper also presents the data-collection platform and pipeline, using cameras above and below a transparent plane followed by Poisson surface reconstruction and voxelization, that produce the dense 3D occupancy labels used to supervise the occupancy predictor.

Load-bearing premise

The load-bearing premise is that the 3D occupancy labels used to supervise the occupancy predictor and dynamics model are accurate, specifically that Poisson surface reconstruction, an algorithm that builds a watertight mesh from a point cloud, captures the plasticine's shape faithfully at every manipulation step; the authors themselves note that it struggles with fine topological changes such as splits and merges.

Editorial extensions

If this is right

  • If the framework works as claimed, a deformable-object manipulation system can be built without real-time mesh reconstruction or particle-based simulation, using only RGB images and a learned occupancy representation.
  • The reported gains over surface- and crop-based representations suggest that internal volumetric information helps learned dynamics models predict plastic deformation, so other volumetric manipulation tasks may benefit from dense occupancy states.
  • Because the shape-based action initialization reportedly stabilizes the irreversible deformation process, the same initialization idea could be applied to other MPC-based deformable-object planners.
  • The transparent-plane data collection pipeline, if accurate, provides a way to generate dense occupancy labels for training in other lab settings.

Reading between the lines

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

  • We would infer that the framework's performance is tied to the transparent-plane setup: without bottom views, the occupancy labels degrade, so the claimed 'given goal shape' generality is most safely read as applying to instrumented environments with full spatial coverage.
  • Because the authors note that Poisson surface reconstruction misses fine topological changes such as splits and merges, goal shapes that require separating or merging material are a plausible failure boundary for the current method.
  • The shape-based initialization could be tested as a standalone module: pairing it with a weaker dynamics model would reveal how much of the planning success comes from initialization versus learned prediction.
  • The occupancy representation may extend beyond plasticine to dough, clay, or foam-like materials, but the sim-to-real gap reported here suggests each new material needs fine-tuning on real data.
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

5 major / 6 minor

Summary. The paper proposes a complete pipeline for manipulating elasto-plastic objects (plasticine) using a dense 3D occupancy representation. A multi-view RGB input is mapped to 3D occupancy by a prediction network; a 3D CNN-GNN model learns the quasi-static transition dynamics on a down-sampled voxel graph; and a model predictive controller with a shape-based action initialization plans pinch actions. The authors contribute a data-collection platform with a transparent plane for full spatial coverage and a pipeline that converts multi-view RGB-D into voxelized occupancy ground truth. Experiments in simulation and on a UR5e platform evaluate the representation, dynamics model, ablations, and several goal shapes (X, T, K, Y, M, two-layer stamps), with comparisons against RoboCraft and RoboCook and against human demonstrations.

Significance. If the reported claims are appropriately scoped, the framework is a useful step: the 3D occupancy representation and the 3D CNN-GNN dynamics model are evaluated with multiple geometric metrics, and the ablation studies isolate the contribution of the 3D CNN, the DCD loss, and the shape-based initialization. The transparent-plane data collection platform is a practical contribution for acquiring full spatial information. The main weaknesses are the mismatch between the universal 'given goal shape' claim and the acknowledged inability to represent topological changes, the absence of quantitative real-world evaluation, and the insufficiently controlled baseline comparison. The paper would be acceptable after these issues are addressed.

major comments (5)
  1. [Section VI, Abstract] The abstract and introduction state that the framework 'can successfully shape the elasto-plastic objects into a given goal shape,' but Section VI concedes that Poisson surface reconstruction (used to generate the occupancy ground truth in Section IV.D) 'struggles to capture fine topological changes such as splits and merges.' All demonstrated goals (X, T, K, Y, M, two-layer stamps) are reachable by pinching without any split or merge. Since the occupancy labels and the learned dynamics cannot represent topological transitions, the universal claim is not supported. Please either demonstrate a task that requires a split or merge, or explicitly restrict the claim to topology-preserving deformations.
  2. [Section V.B.3, Figures 9-12] The real-world manipulation results are presented only as photographs and qualitative statements such as 'closely resembling the goal shape.' There are no quantitative metrics (e.g., EMD, DCD, CD) computed on the real-world final states, no repeated trials with error bars, and no defined success criterion. To substantiate the claim of real-world verification, please provide objective measurements of the final shapes against the goal geometry, with mean and standard deviation over multiple trials.
  3. [Section V.B.2, Table II] The comparison with RoboCraft and RoboCook is not fully controlled. The baselines are trained for 100 epochs while the proposed model is trained for 24 epochs, and no significance tests are reported. The reported error bars overlap substantially (e.g., EMD for RoboCraft is 24.3±2.82 vs. 22.8±2.30 for ours; DCD is 27.3±2.30 vs. 25.3±1.52). Please match the training budget or explain the difference, and add significance testing (e.g., paired bootstrap or t-test) before claiming consistent superiority.
  4. [Section III.D, Section V.A] The model predictive control component is underspecified. The time horizon t, the number of L-BFGS iterations, the line-search or step-size settings, and the exact values of the parameter m in the shape-based initialization are not reported. Since MPC and the initialization module are central contributions, please provide an algorithm pseudocode and a parameter table to make the method reproducible.
  5. [Section IV.D, Section VI] The ground-truth occupancy is generated by Poisson surface reconstruction of RGB-D point clouds, and the authors acknowledge that this 'struggles to capture fine topological changes such as splits and merges.' Because both the occupancy prediction network and the dynamics model are trained on these labels, reconstruction errors (e.g., at thin structures, contacts, or near the transparent plane) propagate into the learned models. Please quantify the reconstruction accuracy against simulator ground truth where available and discuss the sensitivity of the downstream dynamics and MPC to these errors.
minor comments (6)
  1. [Equation (4)] The DCD formula uses undefined notation n_lambda_hat_y and n_lambda_hat_x; please define these quantities explicitly.
  2. [Section V.B.1] 'Averaged over 160 frames (i.e., four pinches)' is ambiguous; please specify the number of frames per pinch and how the 70/20 episode split is used.
  3. [Figure 8] The embedded quantitative table is difficult to read and appears to have rows and columns that may be misaligned; please enlarge the table and ensure the method labels match the legend.
  4. [Figure 12] The label 'Humancraft1' should be 'Human craft 1'.
  5. [Section V.A] Please clarify whether the real-world fine-tuning uses the same occupancy loss as the simulation training and how many real episodes are used for fine-tuning versus evaluation.
  6. [Section III.A] The phrase 'parallelled 2-finger gripper' should be 'parallel 2-finger gripper'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: learned components are evaluated on held-out data and against independent baselines, and the only self-citation is non-load-bearing.

full rationale

The derivation chain is self-contained. The 3D occupancy prediction network (Eq. 2) is supervised by the generated occupancy ground truth of Sec. IV, and Table I evaluates it on held-out simulator frames against that same ground truth; this is a standard supervised-learning protocol, not a fitted input renamed as a prediction. The dynamics model (Eq. 1, Sec. III-C) is trained with EMD/DCD/CD losses on collected transitions and compared against independent RoboCraft and RoboCook baselines (Table II); the metrics are externally computed geometric distances on held-out states, so no fitted parameter is reused as a prediction. The MPC objective in Sec. III-D uses the same loss as the dynamics training loss, but that is a control objective, not a circular derivation, and the shape-based action initialization is explicitly ablated in Fig. 13. The only self-citation is [31] (DOFS), used for data-collection platform details in Sec. IV-A, and it is not load-bearing for the central manipulation claim. Section VI's caveat that Poisson surface reconstruction struggles with splits and merges is a data-quality and generality limitation, not a case where an output reduces to an input. No equation or claim reduces by construction to its own inputs.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

No new physical entities are introduced; the framework relies on hand-chosen hyperparameters and domain assumptions about quasi-static behavior, mesh reconstruction fidelity, and action space limits.

free parameters (4)
  • Total loss weights w1, w2, w3 = w1=0.5, w2=0.4, w3=0.1
    Weights for combining EMD, DCD, and CD in the dynamics model training loss (Section V.B.2).
  • DCD temperature alpha and density scale lambda = alpha=500, lambda=0.5 (evaluation); alpha=20, lambda=0.1 (training)
    Hyperparameters of the density-aware Chamfer distance, chosen empirically (Section V.B.2).
  • Downsampled state size k = k=300
    Number of voxels used to represent the plasticine state after FPS (Section III.B).
  • Shape-based initialization region count m = m in [3,6]
    Number of segmentation regions along each axis for the action initialization heuristic (Section III.D.2).
assumptions (5)
  • domain assumption The deformation dynamics of the elasto-plastic object are quasi-static.
    Invoked in the abstract and Section I; the learned dynamics ignore inertial effects, which is valid only for slow manipulation.
  • domain assumption Poisson surface reconstruction from six RGB-D views yields meshes accurate enough for voxel occupancy ground truth.
    Used in Section IV.D to generate all training labels; the conclusion admits fine topological changes are not captured.
  • domain assumption The occupancy prediction network trained with occluded samples can infer occluded regions from four RGB cameras during deployment.
    Claimed in Section III.B; real-world inference relies on this generalization beyond the training setup.
  • domain assumption Gripper actions are limited to translation in x/y/z, rotation about z, and finger distance.
    Defines the action space in Section III.D.1; goals requiring tilting or arbitrary orientations are not reachable.
  • domain assumption The learned dynamics model is differentiable and L-BFGS can optimize the action sequence effectively.
    The MPC in Section III.D uses gradient-based optimization of the learned model; no guarantee of global optimality is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Manipulating Elasto-Plastic Objects With 3D Occupancy and Learning-Based Predictive Control." pith.science (2026). https://pith.science/paper/FSBUMVPD

@misc{pith2026250516249,
  author       = {Pith},
  title        = {Pith review of: Manipulating Elasto-Plastic Objects With 3D Occupancy and Learning-Based Predictive Control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FSBUMVPD}},
  note         = {Machine review of arXiv:2505.16249}
}
read the original abstract

Manipulating elasto-plastic objects remains a significant challenge due to severe self-occlusion, difficulties of representation, and complicated dynamics. This work proposes a novel framework for elasto-plastic object manipulation with a quasi-static assumption for motions, leveraging 3D occupancy to represent such objects, a learned dynamics model trained with 3D occupancy, and a learning-based predictive control algorithm to address these challenges effectively. We build a novel data collection platform to collect full spatial information and propose a pipeline for generating a 3D occupancy dataset. To infer the 3D occupancy during manipulation, an occupancy prediction network is trained with multiple RGB images supervised by the generated dataset. We design a deep neural network empowered by a 3D convolution neural network (CNN) and a graph neural network (GNN) to predict the complex deformation with the inferred 3D occupancy results. A learning-based predictive control algorithm is introduced to plan the robot actions, incorporating a novel shape-based action initialization module specifically designed to improve the planner efficiency. The proposed framework in this paper can successfully shape the elasto-plastic objects into a given goal shape and has been verified in various experiments both in simulation and the real world.

Figures

Figures reproduced from arXiv: 2505.16249 by the authors.

Figure 1
Figure 1. Overview of our framework and results. We predict 3D occupancy from RGB images, learn the dynamics with a deep neural network, and apply learning-based predictive control with shape￾based action initialization to deform the object into a goal shape. self-occlusion. Although the geometric representation like surface-level point clouds can provide rich spatial information, it is sparse and semantics-lacking. In this w… view at source ↗
Figure 2
Figure 2. Inferred 3D occupancy of manipulation scenario in the simulator. (a) RGB observation. (b) Inferred 3D occupancy visualization. The grey, yellow, and dark blue parts represent the gripper, plasticine, and operating plane, respectively. III. APPROACH A. Problem Formulation In this work, we aim to use a paralleled 2-finger gripper to manipulate an elasto-plastic object (i.e., plasticine) into a given goal shape sgoal i… view at source ↗
Figure 3
Figure 3. 3D occupancy prediction framework. We use four cameras to extract multi-scale features, fuse them with 2D-3D spatial attention, and predict 3D occupancy, supervised by the ground truth. where Ot is the 3D occupancy. Its value is between 0 and 1, representing the occupied probability of the voxel grids. Furthermore, we use the farthest point sampling (FPS) algo￾rithm [23] to down-sample the 3D occupancy O into a spar… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Overall structure of our proposed 3D CNN-Based dynamics model. (a) The 3D occupancy of the current step t and previous three steps t − 3 · · ·t − 1 is down-sampled to construct a voxel-based state graph, and node features fj are extracted through a state encoder. (b) T…
Figure 5
Figure 5. Figure 5: Pipeline of our shape-based action initialization for the gripper. The point cloud sp colored blue is the initial state of plasticine and the point cloud sgt colored yellow is the goal of plasticine (i.e., letter “A”). (a) Align sp to sgt. (a)→(b) Filter out points clo…
Figure 7
Figure 7. Figure 7: Pipeline for occupancy ground truth generation. (a) Converting RGB-Ds into dense and well-registered point clouds Praw filtered by M. (b) Point cloud Proi of DO and gripper. (c) Reconstruct deformed meshes M, Mp, Mg1 Mg2 . (d) Dense 3D occupancy ground truth. (e) Down-…
Figure 9
Figure 9. Figure 9: Manipulation results of single-layer objects in real world. Goal State Initial State Manipulation Results 𝑯" 𝟏 𝑯" 𝟐 𝑯" 𝟏 𝑯" 𝟐 Simulation Real-world 𝑯𝟏 𝑯𝟐 𝑯𝟏 𝑯𝟐 𝑯𝟏 𝑯𝟐 = 1.6 𝑯𝟏 𝑯𝟐 = 1.5 1 2 3 4 𝑯" 𝟐 𝑯" 𝟏 𝑯" 𝟏 𝑯" 𝟐 1 2 3 𝑯# 𝟏 𝑯# 𝟐 ≈ 1.5 4 𝑯# 𝟏 𝑯# 𝟐 ≈ 1.6 1.5cm 2.3cm 1.5cm…
Figure 11
Figure 11. Figure 11: Generalization demonstrations in terms of initial shapes, colors, and materials. Six trials are conducted in four initial shapes, three colors, and two materials (dense play-doh vs. spongy form clay). T Y M Ours Human craft 1 Human craft 2 (b) Tools for human craft (a…
Figure 12
Figure 12. Figure 12: Comparison of experimental results. (a) The gripper for robot manipulation. (b) Tool for human manipulation inspired by [35]. (c) The shaping results of our method, two amateur humans, and RobotCraft [4] [PITH_FULL_IMAGE:figures/full_fig_p007_12.png]
Figure 13
Figure 13. Figure 13: Ablation study of shape-based action initialization. “w/o shape-based” means that the actions are randomly initialized during manipulation. (a) Visualization of actions initialization results. (b) The evaluation results using EMD, DCD, and CD after the first action. i…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 26 canonical work pages

  1. [1]

    Learning foresightful dense visual affordance for de- formable object manipulation,

    R. Wuet al., “Learning foresightful dense visual affordance for de- formable object manipulation,” inICCV, pp. 10947–10956, 2023

    work page 2023

  2. [2]

    Foldsformer: Learning sequential multi-step cloth ma- nipulation with space-time attention,

    K. Moet al., “Foldsformer: Learning sequential multi-step cloth ma- nipulation with space-time attention,”RAL, vol. 8, no. 2, pp. 760–767, 2022

    work page 2022

  3. [3]

    Learning-based MPC with safety filter for constrained deformable linear object manipulation,

    Y . Tanget al., “Learning-based MPC with safety filter for constrained deformable linear object manipulation,”RAL, 2024

    work page 2024

  4. [4]

    RoboCraft: Learning to see, simulate, and shape elasto- plastic objects in 3D with graph networks,

    H. Shiet al., “RoboCraft: Learning to see, simulate, and shape elasto- plastic objects in 3D with graph networks,”IJRR, vol. 43, no. 4, pp. 533–549, 2024

    work page 2024

  5. [5]

    DiffSRL: Learning dynamical state representation for deformable object manipulation with differentiable simulation,

    S. Chenet al., “DiffSRL: Learning dynamical state representation for deformable object manipulation with differentiable simulation,”RAL, vol. 7, no. 4, pp. 9533–9540, 2022

    work page 2022

  6. [6]

    Learning closed-loop dough manipulation using a differ- entiable reset module,

    C. Qiet al., “Learning closed-loop dough manipulation using a differ- entiable reset module,”RAL, vol. 7, no. 4, pp. 9857–9864, 2022

    work page 2022

  7. [7]

    6-DoF GraspNet: Variational grasp generation for object manipulation,

    A. Mousavianet al., “6-DoF GraspNet: Variational grasp generation for object manipulation,” inICCV, pp. 2901–2910, 2019

    work page 2019

  8. [8]

    Realtime simulation of thin-shell deformable materials using CNN-based mesh embedding,

    Q. Tanet al., “Realtime simulation of thin-shell deformable materials using CNN-based mesh embedding,”RAL, vol. 5, no. 2, pp. 2325–2332, 2020

    work page 2020

Show all 35 references
  1. [9]

    Sim-to-real reinforcement learning for deformable object manipulation,

    J. Mataset al., “Sim-to-real reinforcement learning for deformable object manipulation,” inCoRL, pp. 734–743, 2018

    2018

  2. [10]

    Learning visual-based deformable object rearrangement with local graph neural networks,

    Y . Denget al., “Learning visual-based deformable object rearrangement with local graph neural networks,”Complex & Intelligent Systems, vol. 9, no. 5, pp. 5923–5936, 2023

    2023

  3. [11]

    Learning particle dynamics for manipulating rigid bodies, deformable objects, and fluids,

    Y . Liet al., “Learning particle dynamics for manipulating rigid bodies, deformable objects, and fluids,”arXiv:1810.01566, 2018

    2018 arXiv

  4. [12]

    DefGoalNet: Contextual goal learning from demonstra- tions for deformable object manipulation,

    B. Thachet al., “DefGoalNet: Contextual goal learning from demonstra- tions for deformable object manipulation,”ICRA, pp. 3145–52, 2024

    2024

  5. [13]

    DeformerNet: Learning bimanual manipulation of 3D deformable objects,

    B. Thachet al., “DeformerNet: Learning bimanual manipulation of 3D deformable objects,”arXiv:2305.04449, 2023

    2023 arXiv

  6. [14]

    SculptDiff: Learning robotic clay sculpting from humans with goal conditioned diffusion policy,

    A. Bartschet al., “SculptDiff: Learning robotic clay sculpting from humans with goal conditioned diffusion policy,” inIROS, pp. 7307– 7314, 2024

    2024

  7. [15]

    DeformNet: Latent space modeling and dynamics predic- tion for deformable object manipulation,

    C. Liet al., “DeformNet: Latent space modeling and dynamics predic- tion for deformable object manipulation,”arXiv:2402.07648, 2024

    2024 arXiv

  8. [16]

    The surprising effectiveness of linear models for visual foresight in object pile manipulation,

    H. J. T. Suhet al., “The surprising effectiveness of linear models for visual foresight in object pile manipulation,” inAlgorithmic Foundations of Robotics XIV, pp. 347–363, 2021

    2021

  9. [17]

    Neural field dynamics model for granular object piles manipulation,

    S. Xueet al., “Neural field dynamics model for granular object piles manipulation,” inCoRL, pp. 2821–2837, 2023

    2023

  10. [18]

    Dynamic-resolution model learning for object pile manipulation,

    Y . Wanget al., “Dynamic-resolution model learning for object pile manipulation,”arXiv:2306.16700, 2023

    2023 arXiv

  11. [19]

    Transporter networks: Rearranging the visual world for robotic manipulation,

    A. Zenget al., “Transporter networks: Rearranging the visual world for robotic manipulation,” inCoRL, pp. 726–747, 2021

    2021

  12. [20]

    Learning compositional Koopman operators for model- based control,

    Y . Liet al., “Learning compositional Koopman operators for model- based control,”arXiv:1910.08264, 2019

    1910 arXiv

  13. [21]

    RoboCook: Long-horizon elasto-plastic object manipula- tion with diverse tools,

    H. Shiet al., “RoboCook: Long-horizon elasto-plastic object manipula- tion with diverse tools,”arXiv:2306.14447, 2023

    2023 arXiv

  14. [22]

    Sculptbot: Pre-trained models for 3D deformable object manipulation,

    A. Bartschet al., “Sculptbot: Pre-trained models for 3D deformable object manipulation,” inICRA, pp. 12548–12555, 2024

    2024

  15. [23]

    Fast marching farthest point sampling,

    C. Moenninget al., “Fast marching farthest point sampling,” Tech. Rep., University of Cambridge, Computer Laboratory, 2003

    2003

  16. [24]

    Deep residual learning for image recognition,

    K. Heet al., “Deep residual learning for image recognition,” inCVPR, pp. 770–778, 2016

    2016

  17. [25]

    SurroundOcc: Multi-camera 3D occupancy prediction for autonomous driving,

    Y . Weiet al., “SurroundOcc: Multi-camera 3D occupancy prediction for autonomous driving,” inICCV, pp. 21729–21740, 2023

    2023

  18. [26]

    PV-RCNN: Point-voxel feature set abstraction for 3D object detection,

    S. Shiet al., “PV-RCNN: Point-voxel feature set abstraction for 3D object detection,” inCVPR, pp. 10529–10538, 2020

    2020

  19. [27]

    A metric for distributions with applications to image databases,

    Y . Rubneret al., “A metric for distributions with applications to image databases,” inICCV, pp. 59–66, 1998

    1998

  20. [28]

    Density-aware Chamfer Distance as a comprehensive metric for point cloud completion,

    T. Wuet al., “Density-aware Chamfer Distance as a comprehensive metric for point cloud completion,”arXiv:2111.12702, 2021

    2021 arXiv

  21. [29]

    Practical Methods of Optimization,

    R. Fletcher, “Practical Methods of Optimization,” John Wiley & Sons, 2013

    2013

  22. [30]

    PokeFlex: Towards a real-world dataset of deformable objects for robotic manipulation,

    J. Obristet al., “PokeFlex: Towards a real-world dataset of deformable objects for robotic manipulation,”arXiv:2409.17124, 2024

    2024 arXiv

  23. [31]

    DOFS: A real-world 3D deformable object dataset with full spatial information for dynamics model learning,

    Z. Zhanget al., “DOFS: A real-world 3D deformable object dataset with full spatial information for dynamics model learning,” arXiv:2410.21758, 2024

    2024 arXiv

  24. [32]

    A density-based algorithm for discovering clusters in large spatial databases with noise,

    M. Esteret al., “A density-based algorithm for discovering clusters in large spatial databases with noise,” inProceedings of KDD, vol. 96, no. 34, pp. 226–231, 1996

    1996

  25. [33]

    Poisson surface reconstruction,

    M. Kazhdanet al., “Poisson surface reconstruction,” inProc. Eurograph- ics Symp. Geometry Processing (SGP), vol. 7, no. 4, 2006

    2006

  26. [34]

    PlasticineLab: A soft-body manipulation benchmark with differentiable physics,

    Z. Huanget al., “PlasticineLab: A soft-body manipulation benchmark with differentiable physics,”arXiv:2104.03311, 2021

    2021 arXiv

  27. [35]

    Universal Manipulation Interface: In-the-wild robot teaching without in-the-wild robots,

    C. Chiet al., “Universal Manipulation Interface: In-the-wild robot teaching without in-the-wild robots,”arXiv:2402.10329, 2024

    2024 arXiv

Pith tools

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