Pith. sign in

REVIEW 3 major objections 5 minor 31 references

PinchBot: Long-Horizon Deformable Manipulation with Guided Diffusion Policy

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read A single goal-conditioned diffusion policy can sculpt clay bowls of three sizes using only pinches.

desk verdict A useful new long-horizon benchmark with a solid ablation, but the three-run statistics don't support the 'across all metrics' claim and the collision projection hides a radial-symmetry assumption. read the letter →

arxiv 2507.17846 v1 pith:LWBFJ42L submitted 2025-07-23 cs.RO

classification cs.RO
keywords diffusionpolicygoal-conditioneddeformablemanipulationclaysculptingpinchpotterypointcloudembeddingtaskprogresspredictioncollision-constrainedprojection
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

PinchBot sets out to show that freehand pinch pottery, a long-horizon deformable manipulation task requiring 21 to 31 pinches per object, can be handled by a single imitation-learning policy rather than a per-goal specialist or a hand-built planner. It is a goal-conditioned diffusion policy that takes point clouds of the current clay and the target bowl, embeds them with a pre-trained 3D point cloud model, and denoises an action chunk while also predicting how far along the demonstration the current state is. A collision-projection step then pushes any generated pinch position that would land inside the clay wall out to the wall boundary. The paper reports that the same trained policy adapts its sculpting sequence to create 8 cm, 10 cm, and 12 cm diameter bowls from clay cylinders, using only pinch actions and stopping on its own through a learned termination signal. The authors frame the task as a tougher benchmark than the alphabet-shaped clay goals used in prior work and note that the behavior-cloning framework remains limited in generalizability.

What carries the argument

The mechanism that carries PinchBot is goal-conditioned action denoising augmented by three guidance signals. A pretrained point-cloud embedding compresses the current clay and goal bowl into a shared latent space, so the policy can condition on goal shape without retraining per goal; pretraining on a large reconstruction corpus compensates for the small real-world demonstration set. A task-progress head predicts a continuous value from -1 to 1 indicating where the current state falls along the demonstration, giving the policy a learned sense of how much work remains and when to emit the final-action flag. Collision-constrained projection fits a circle in the horizontal plane that contains more than 95% of the clay points and projects any generated pinch position inside that circle out to its edge, preserving height, orientation, and fingertip gap. The paper's key simplification is that the clay cross-section is treated as circular at every step of the long sequence.

What would settle it

Run the full PinchBot policy on a 10 cm goal and record the ratio of the largest to the smallest radial distance of the clay point cloud in the horizontal plane before every pinch. If the circle-fitting collision projection is doing the work, states with a ratio meaningfully above 1 (an asymmetric or uneven wall) should show pinch positions clipped inside or outside the actual wall, visible as wall punctures or sharply higher Chamfer distance; observing no such failures at high asymmetry would falsify the claim that the radial-circle simplification is the key safety mechanism.

Watch

Extended reading notes

Core claim

The paper's central claim is that one goal-conditioned diffusion policy, trained on a small set of kinesthetic demonstrations rotated to 3600 trajectories, can produce a family of pottery bowls with no tools and no wheel. The learned policy embeds the current and goal point clouds with a pretrained encoder, conditions action denoising on that latent plus the previous action and a progress indicator, and executes pinches in chunks of four before replanning. The collision-projection module fits a circle to the horizontal projection of the clay cloud and moves any unsafe pinch position to the circle boundary, which is what allows the policy to avoid irrecoverable wall-piercing errors over dozens of steps. In the reported experiments, policies using the transformer-based pretrained embedding match goal point clouds with lower Chamfer and Earth Mover distances than the point-encoder baseline, reach the final bowl in fewer actions, and lose performance when pretraining, collision projection, or the diffusion objective is removed.

Load-bearing premise

The load-bearing assumption is that the clay pot can always be described by a circle in the horizontal plane, found by fitting a diameter that contains more than 95% of the projected points; if the pot is asymmetric or the wall is not radially symmetric, projecting the gripper onto that circle can push it into the clay or miss the wall entirely.

Editorial extensions

If this is right

  • A single policy can serve multiple target shapes: the goal point cloud alone steers both where to pinch and when to stop, so no per-diameter training is needed.
  • Autonomous stop decisions can be learned rather than set by threshold, since the progress and termination head tells the policy when the goal is reached.
  • Long-horizon deformable tasks expose differences between embedding models and guidance schemes that short shaping tasks do not reveal, making them useful benchmarks for imitation learning.
  • Collision projection turns a policy that occasionally proposes harmful pinches into one that can run for tens of actions without catastrophic wall damage.
  • Pre-training the state and goal encoder on a large 3D reconstruction task is what makes a 20-demonstration dataset sufficient for this class of policy.

Reading between the lines

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

  • If the circular cross-section assumption were replaced by an ellipse or a local wall model, the same collision-projection idea would likely extend to asymmetric pots and non-circular rims; this is a natural next test the paper does not run.
  • The task-progress signal may transfer to other variable-length deformable tasks that have no explicit subtask structure, because it only requires knowing roughly how far along the demonstration the current state is.
  • Because the policy generates future actions in chunks, combining progress conditioning with latent-space steering could enlarge the range of achievable diameters without collecting new demonstrations, an option the authors themselves point to.
  • The reported failure of sub-goal conditioning on the 12 cm goal suggests that intermediate-goal guidance can confuse a policy when intermediate states of one goal resemble final states of another; checking this in simpler tasks could sharpen the design of sub-goal methods.
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 / 5 minor

Summary. The paper introduces PinchBot, a goal-conditioned diffusion policy for long-horizon pinch-based pottery shaping on a Franka robot. The method combines pre-trained 3D point cloud embeddings (PointBERT and a DP3-style PointNet), task-progress and sub-goal guidance, and a collision-constrained action projection. The authors collect 20 real-world kinesthetic demonstrations, augment them by 2-degree rotations to 3600 trajectories, and train a single policy to produce 8, 10, and 12 cm diameter bowls from clay cylinders. Evaluation reports Chamfer Distance, Earth Mover's Distance, and diameter MSE over three real-world runs per variant, alongside ablations and a T-SNE latent-space analysis. The paper claims PointBERT variants outperform DP3 PointNet variants across all metrics, that task-progress guidance improves behavior, and that pre-training, collision projection, and diffusion-style training all contribute to success.

Significance. If the results hold, PinchBot is a meaningful step for long-horizon deformable manipulation: it is one of few systems that learns a single goal-conditioned policy for a 21-31 action pinch-pottery task, and it provides a public demonstration dataset. The systematic comparison of point cloud embedders and guidance variants, plus the ablation study, are useful for the community. The authors are also honest about generalizability limits. However, the strength of the comparative claims currently exceeds the statistical support, and the collision-projection mechanism relies on an unverified radial-symmetry assumption for exactly the slanted-wall goals the system is asked to produce.

major comments (3)
  1. [Section V-B / Table I] The claim that "PointBERT policies outperform the DP3 PointNet variants across all quantitative metrics" is not supported by Table I. For example, on the 8 cm goal with Continuous Guidance, PointBERT has CD 9.9 ± 0.5 mm versus DP3 PointNet's 8.5 ± 0.9 mm, and on the 10 cm goal with Continuous Guidance, PointBERT has CD 7.3 ± 0.2 mm versus 7.0 ± 0.2 mm. Even where PointBERT appears better, most differences are within one standard deviation (e.g., 10 cm Binary CD 7.1 ± 0.2 vs 7.2 ± 0.3; 10 cm EMD 6.3 ± 0.5 vs 6.5 ± 0.4). With only three runs per variant, these overlaps are not adequately discussed. Please temper the claim to a per-condition, uncertainty-aware comparison, and consider reporting paired differences or effect sizes.
  2. [Section IV-E] The collision projection assumes the clay is radially symmetric about the fitted circle's center and that the circle radius is the wall radius at every z. This assumption is violated by the paper's own target set: Section III-D states that larger bowl diameters are correlated with angled walls, and the 12 cm goal is explicitly slanted. For a slanted wall, preserving z while projecting (x,y) onto the fitted circle can move the gripper into the clay or away from the wall. Asymmetric dents created by pinch errors can also bias the circle fit. Since Table III shows that removing projection degrades results and Section IV-E states that a single bad pinch can be irrecoverable, this assumption is load-bearing. Please add a quantitative safeguards check during rollouts: verify that each projected action position is outside the actual clay point cloud at the action's z (e.g., via nearest-neighbor distance in the horizontal plane), report any violations, and consider using a conservative local wall estimate instead of a global circle.
  3. [Section V-A] The diameter MSE metric is not defined precisely enough to be reproducible. Please specify how the final pottery diameter is measured from the final point cloud (e.g., circle fit on the top rim, maximum horizontal extent at a given height, or manual measurement), and whether the same procedure is applied to the goal shape.
minor comments (5)
  1. [Section IV-E] The circle-fitting and projection procedure is described in words only; a formal algorithm or pseudocode with the exact point-selection criterion and projection rule would improve reproducibility.
  2. [Section VI] There are several typos, including "emebeddings" (Section V-E), "wheras" (Figure 6 caption), and "subtantially" (Section VI). Please proofread.
  3. [Section IV-D] Please clarify how the predicted continuous task-progress value is used at inference time: is it part of the action output, used for termination, or only an auxiliary training target? The text does not make this operational role explicit.
  4. [Section IV-A] The paper reports adding a 3-layer MLP head to project each embedding to size 512, but does not give training hyperparameters for the pre-training or fine-tuning stages. Reporting learning rate, batch size, training steps, and diffusion settings would improve reproducibility, even if some details are deferred to the project website.
  5. [Section III-D] The rotation augmentation is described as producing 3600 trajectories from 20 demonstrations; please state explicitly that this assumes 180 rotations per demonstration (2-degree increments), which appears to be the case but is not stated.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: PinchBot's central claims are validated by real-robot experiments and ablations, not by self-referential fits.

full rationale

PinchBot is an empirical systems paper. Its central claim—that a single goal-conditioned diffusion policy plus pretrained point-cloud embeddings, task-progress guidance, and collision-constrained projection can create 8/10/12 cm bowls—is supported by physical robot rollouts (three runs per variant, Table I), ablations (Table III), and standard point-cloud metrics (CD, EMD, MSE). No derivation in the paper reduces to its inputs by construction. The task-progress label γ is a training target computed from demonstration trajectory index; it is not a fitted parameter later renamed as a prediction. The collision projection (Sec. IV-E) is a heuristic post-processor that fits a circle to >95% of the xy-projected clay points and pushes unsafe actions to that circle's edge; while this assumes radial symmetry and is a legitimate correctness risk (unsupported for slanted/asymmetric walls), it is not circular because its contribution is established by an ablation against real-robot outcomes rather than by construction from the goal. Self-citations to SculptDiff [11], SculptBot [10], and LLM-Craft [12] appear as contextual prior work and as sources of pretraining/architecture choices; they are not load-bearing evidence for the paper's own empirical claims, and the pretraining contribution is additionally tested in the ablation ('No pretraining' rows in Table III). The limitations paragraph in Sec. VI explicitly acknowledges limited generalizability, which is an honest scope statement rather than an admission of circularity. Overall, no circular step meets the evidentiary bar of Eq. X = Eq. Y by construction or fitted-parameter-renamed-as-prediction.

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

The central claims rest on standard assumptions of imitation learning: the demonstrated trajectories are representative, the point cloud perception is accurate, ShapeNet pre-training transfers, and the diffusion policy can model multimodal action distributions. No new physical entities are introduced.

free parameters (4)
  • Circle-fit threshold = 0.95
    In collision projection, a circle is fit to the current clay point cloud by finding the diameter that fits more than 95% of points; this hand-chosen threshold determines whether an action is projected.
  • Rotation augmentation increment = 2 degrees
    Demonstrations are rotated about the z-axis in 2 degree increments to expand 20 demos to 3600 trajectories; this choice affects policy diversity and smoothness.
  • Action prediction and execution horizons = 16 and 4
    Diffusion policy plans 16 actions and replans every 4 actions at test time; these horizons are chosen without an ablation and affect long-horizon behavior.
  • Sub-goal step size = 8
    For sub-goal variants, the number of sub-goals N is determined by prediction horizon divided by step size, i.e., N=2; this choice affects the sub-goal guidance.
assumptions (4)
  • domain assumption The 20 kinesthetic demonstrations, augmented by rotation, are representative enough to train a policy that generalizes across clay cylinder sizes and goal diameters.
    The policy is trained on 20 real demos plus rotations; the paper assumes this is sufficient for modest generalization, and the limitations section admits limited generalizability.
  • domain assumption The 5-camera point cloud fusion with ICP alignment produces an accurate and consistent state estimate of the clay.
    Perception quality is load-bearing for behavior cloning; the paper does not quantify point cloud reconstruction error.
  • domain assumption Pre-training point cloud encoders on ShapeNet reconstruction transfers to clay point clouds.
    The method relies on ShapeNet pre-training to overcome the small dataset; this transfer assumption is investigated only indirectly through ablation results.
  • standard math Diffusion policy training can model the multimodal distribution of valid pinch action sequences.
    The paper adopts the diffusion policy framework from prior work, assuming its denoising process can represent the multiple valid action sequences for pinch pottery.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PinchBot: Long-Horizon Deformable Manipulation with Guided Diffusion Policy." pith.science (2026). https://pith.science/paper/LWBFJ42L

@misc{pith2026250717846,
  author       = {Pith},
  title        = {Pith review of: PinchBot: Long-Horizon Deformable Manipulation with Guided Diffusion Policy},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LWBFJ42L}},
  note         = {Machine review of arXiv:2507.17846}
}
read the original abstract

Pottery creation is a complicated art form that requires dexterous, precise and delicate actions to slowly morph a block of clay to a meaningful, and often useful 3D goal shape. In this work, we aim to create a robotic system that can create simple pottery goals with only pinch-based actions. This pinch pottery task allows us to explore the challenges of a highly multi-modal and long-horizon deformable manipulation task. To this end, we present PinchBot, a goal-conditioned diffusion policy model that when combined with pre-trained 3D point cloud embeddings, task progress prediction and collision-constrained action projection, is able to successfully create a variety of simple pottery goals. For experimental videos and access to the demonstration dataset, please visit our project website: https://sites.google.com/andrew.cmu.edu/pinchbot/home.

Figures

Figures reproduced from arXiv: 2507.17846 by the authors.

Figure 1
Figure 1. A full pottery sculpting sequence of our trained PinchBot policy. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The hardware setup. a) The workspace for our experiments consisting of 4 Intel RealSense D415 cameras, and an elevated stage the clay remains [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. The PinchBot method. a) Pre-training embeddings for the reconstruction task. b) Collecting 20 real-world pinch pottery demonstrations with kinesthetic [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: The final shapes conditioned on the 8cm, 10cm and 12cm bowl diameter goals across point cloud embedding and policy variant. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: The final shapes for the ablation study variants exploring how pre-training, collision projection and diffusion policy-style training impact the final [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: T-SNE plots of the embeddings of the rotation augmented demonstration trajectories across model variations. The colorbar variation along the x [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 20 canonical work pages

  1. [1]

    Point cloud matters: Rethinking the impact of different observation spaces on robot learning,

    H. Zhu, Y . Wang, D. Huang, W. Ye, W. Ouyang, and T. He, “Point cloud matters: Rethinking the impact of different observation spaces on robot learning,” Advances in Neural Information Processing Systems , vol. 37, pp. 77 799–77 830, 2024

  2. [2]

    Differentiable Particles for General-Purpose Deformable Object Manipulation

    S. Chen, Y . Xu, C. Yu, L. Li, and D. Hsu, “Differentiable particles for general-purpose deformable object manipulation,” arXiv preprint arXiv:2405.01044, 2024

  3. [3]

    General-purpose clothes manipulation with semantic keypoints,

    Y . Deng and D. Hsu, “General-purpose clothes manipulation with semantic keypoints,” arXiv preprint arXiv:2408.08160 , 2024

  4. [4]

    Ssfold: Learning to fold arbitrary crumpled cloth using graph dynamics from human demonstration,

    C. Zhou, H. Xu, J. Hu, F. Luan, Z. Wang, Y . Dong, Y . Zhou, and B. He, “Ssfold: Learning to fold arbitrary crumpled cloth using graph dynamics from human demonstration,” IEEE Transactions on Automation Science and Engineering, 2025

  5. [5]

    Diffusion dynamics models with generative state estimation for cloth manipulation,

    T. Tian, H. Li, B. Ai, X. Yuan, Z. Huang, and H. Su, “Diffusion dynamics models with generative state estimation for cloth manipulation,” arXiv preprint arXiv:2503.11999, 2025

  6. [6]

    Dexdlo: Learning goal-conditioned dexterous policy for dynamic manipulation of deformable linear objects,

    S. Zhaole, J. Zhu, and R. B. Fisher, “Dexdlo: Learning goal-conditioned dexterous policy for dynamic manipulation of deformable linear objects,” in 2024 IEEE international conference on robotics and automation (ICRA). IEEE, 2024, pp. 16 009–16 015

  7. [7]

    Robotic cutting of fruits and vegetables: Modeling the effects of deformation, fracture toughness, knife edge geometry, and motion,

    P. Jamdagni and Y .-B. Jia, “Robotic cutting of fruits and vegetables: Modeling the effects of deformation, fracture toughness, knife edge geometry, and motion,” IEEE Transactions on Robotics , 2024

  8. [8]

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

    H. Shi, H. Xu, Z. Huang, Y . Li, and J. Wu, “Robocraft: Learning to see, simulate, and shape elasto-plastic objects in 3d with graph networks,” The International Journal of Robotics Research, vol. 43, no. 4, pp. 533– 549, 2024

Show all 31 references
  1. [9]

    Robocook: Long-horizon elasto-plastic object manipulation with diverse tools,

    H. Shi, H. Xu, S. Clarke, Y . Li, and J. Wu, “Robocook: Long-horizon elasto-plastic object manipulation with diverse tools,” arXiv preprint arXiv:2306.14447, 2023

  2. [10]

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

    A. Bartsch, C. Avra, and A. B. Farimani, “Sculptbot: Pre-trained models for 3d deformable object manipulation,” in 2024 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2024, pp. 12 548–12 555

  3. [11]

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

    A. Bartsch, A. Car, C. Avra, and A. B. Farimani, “Sculptdiff: Learning robotic clay sculpting from humans with goal conditioned diffusion policy,” in 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . IEEE, 2024, pp. 7307–7314

  4. [12]

    Llm-craft: Robotic crafting of elasto-plastic objects with large language models,

    A. Bartsch and A. B. Farimani, “Llm-craft: Robotic crafting of elasto-plastic objects with large language models,” arXiv preprint arXiv:2406.08648, 2024

  5. [13]

    Manipulating elasto-plastic objects with 3d occupancy and learning- based predictive control,

    Z. Zhang, X. Chu, Y . Tang, L. Zhao, J. Huang, Z. Jiang, and K. S. Au, “Manipulating elasto-plastic objects with 3d occupancy and learning- based predictive control,” IEEE Robotics and Automation Letters , 2025

  6. [14]

    Learning closed-loop dough manipulation using a differentiable reset module,

    C. Qi, X. Lin, and D. Held, “Learning closed-loop dough manipulation using a differentiable reset module,” IEEE Robotics and Automation Letters, vol. 7, no. 4, pp. 9857–9864, 2022

  7. [15]

    Planning with spatial-temporal abstraction from point clouds for deformable object manipulation,

    X. Lin, C. Qi, Y . Zhang, Z. Huang, K. Fragkiadaki, Y . Li, C. Gan, and D. Held, “Planning with spatial-temporal abstraction from point clouds for deformable object manipulation,” CoRL, 2022

  8. [16]

    Doughnet: A visual predictive model for topological manipulation of deformable objects,

    D. Bauer, Z. Xu, and S. Song, “Doughnet: A visual predictive model for topological manipulation of deformable objects,” in European Con- ference on Computer Vision . Springer, 2024, pp. 92–108

  9. [17]

    Dexdeform: Dexterous deformable object manipulation with human demonstrations and differentiable physics,

    S. Li, Z. Huang, T. Chen, T. Du, H. Su, J. B. Tenenbaum, and C. Gan, “Dexdeform: Dexterous deformable object manipulation with human demonstrations and differentiable physics,” arXiv preprint arXiv:2304.03223, 2023

  10. [18]

    Make a donut: Hierarchical emd-space planning for zero-shot de- formable manipulation with tools,

    Y . You, B. Shen, C. Deng, H. Geng, S. Wei, H. Wang, and L. Guibas, “Make a donut: Hierarchical emd-space planning for zero-shot de- formable manipulation with tools,” IEEE Robotics and Automation Letters, 2025

  11. [19]

    Robotic dough shaping,

    J. Ondras, D. Ni, X. Deng, Z. Gu, H. Zheng, and T. Bhattacharjee, “Robotic dough shaping,” in 2022 22nd International Conference on Control, Automation and Systems (ICCAS) . IEEE, 2022, pp. 300–307

  12. [20]

    Planning framework for robotic pizza dough stretching with a rolling pin,

    J.-T. Kim, F. Ruggiero, V . Lippiello, and B. Siciliano, “Planning framework for robotic pizza dough stretching with a rolling pin,” in Robot dynamic manipulation: perception of deformable objects and nonprehensile manipulation control. Springer, 2022, pp. 229–253

  13. [21]

    Diffusion policy: Visuomotor policy learning via ac- tion diffusion,

    C. Chi, Z. Xu, S. Feng, E. Cousineau, Y . Du, B. Burchfiel, R. Tedrake, and S. Song, “Diffusion policy: Visuomotor policy learning via ac- tion diffusion,” The International Journal of Robotics Research , p. 02783649241273668, 2023

  14. [22]

    3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations,

    Y . Ze, G. Zhang, K. Zhang, C. Hu, M. Wang, and H. Xu, “3d diffusion policy: Generalizable visuomotor policy learning via simple 3d representations,” arXiv preprint arXiv:2403.03954 , 2024

  15. [23]

    Ropotter: Toward robotic pottery and deformable object manipulation with structural priors,

    U. Yoo, A. Hung, J. Francis, J. Oh, and J. Ichnowski, “Ropotter: Toward robotic pottery and deformable object manipulation with structural priors,” in 2024 IEEE-RAS 23rd International Conference on Humanoid Robots (Humanoids). IEEE, 2024, pp. 843–850

  16. [24]

    Learning fine- grained bimanual manipulation with low-cost hardware,

    T. Z. Zhao, V . Kumar, S. Levine, and C. Finn, “Learning fine- grained bimanual manipulation with low-cost hardware,” arXiv preprint arXiv:2304.13705, 2023

  17. [25]

    Point-bert: Pre-training 3d point cloud transformers with masked point modeling,

    X. Yu, L. Tang, Y . Rao, T. Huang, J. Zhou, and J. Lu, “Point-bert: Pre-training 3d point cloud transformers with masked point modeling,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 19 313–19 322

  18. [26]

    Shapenet: An information- rich 3d model repository,

    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

  19. [27]

    Pointnet: Deep learning on point sets for 3d classification and segmentation,

    C. R. Qi, H. Su, K. Mo, and L. J. Guibas, “Pointnet: Deep learning on point sets for 3d classification and segmentation,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2017, pp. 652–660

  20. [28]

    Planning and reasoning with 3d deformable objects for hierarchical text-to-3d robotic shaping,

    A. Bartsch and A. B. Farimani, “Planning and reasoning with 3d deformable objects for hierarchical text-to-3d robotic shaping,” IEEE Robotics and Automation Letters , pp. 1–8, 2025

  21. [29]

    The earth mover’s distance as a metric for image retrieval,

    Y . Rubner, C. Tomasi, and L. J. Guibas, “The earth mover’s distance as a metric for image retrieval,” International journal of computer vision , vol. 40, pp. 99–121, 2000

  22. [30]

    A point set generation network for 3d object reconstruction from a single image,

    H. Fan, H. Su, and L. J. Guibas, “A point set generation network for 3d object reconstruction from a single image,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2017, pp. 605–613

  23. [31]

    Steering your diffusion policy with latent space reinforcement learning,

    A. Wagenmaker, M. Nakamoto, Y . Zhang, S. Park, W. Yagoub, A. Naga- bandi, A. Gupta, and S. Levine, “Steering your diffusion policy with latent space reinforcement learning,” arXiv preprint arXiv:2506.15799 , 2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.