Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

ArtGS:3D Gaussian Splatting for Interactive Visual-Physical Modeling and Manipulation of Articulated Objects

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read ArtGS proposes that a robot can learn an articulated object's joints by reconstructing it as 3D Gaussians, touching it, and optimizing rendered motion against observed RGB-D frames, and reports the lowest joint errors and highest…

desk verdict A genuinely new closed-loop articulation refinement idea, let down by messy tables and an unevaluated VLM discrete-type assumption. read the letter →

arxiv 2507.02600 v1 pith:W3UHNMOW submitted 2025-07-03 cs.RO

Qiaojun Yu , Xibin Yuan , Yu jiang , Junting Chen , Dongzhe Zheng , Ce Hao , Yang You , Yixing Chen
show 3 more authors
Yao Mu Liu Liu Cewu Lu
This is my paper
classification cs.RO
keywords 3DGaussianSplattingarticulatedobjectsjointparameterestimationroboticmanipulationvision-languagemodeldifferentiablerenderinglinearblendskinningdigitaltwin
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 claims that a robot can build an accurate working model of an articulated object—where its joints are, what type they are, and how they move—by reconstructing the object as 3D Gaussians, asking a vision-language model for initial joint guesses, and then refining those guesses through its own manipulation. The refinement renders the moving object with differentiable Gaussian splatting and adjusts joint parameters until the rendered frames match what the cameras record while the robot touches the object. This closes a loop that point-cloud-based estimators lack: observe, touch, compare, correct. The paper reports the lowest joint-axis and joint-origin errors among the compared methods, and manipulation success rates of 62–90 percent in simulation and 7–10 out of 10 trials on real objects.

What carries the argument

The load-bearing object is the rigid skeleton $B(\theta) = (B_0, B_1, \dots, B_K)$, a list of $SE(3)$ bone transformations that map canonical Gaussian coordinates to an articulated pose; $B_0$ is the identity and each $B_i$ is either a revolute rotation built from the Rodrigues formula or a prismatic translation along the joint axis. Gaussian points are attached to the bones by learnable skinning weights $W$ and deformed by linear blend skinning, so the rendered image is differentiable with respect to the joint parameters themselves. A fine-tuned vision-language model supplies the initial bone guesses from a synthesized frontal view, and the parameters are then optimized against the multi-view RGB-D rendering loss of Eq. 9. On the robot side, Modified Denavit-Hartenberg forward kinematics reconstructs each arm link as its own set of Gaussians, which is what gives the method its cross-embodiment property.

What would settle it

Force the vision-language model to output a wrong joint type for a known object—for example, labeling a prismatic drawer as revolute—and run the full optimization; if the rendered-image loss still drives the joint parameters to the true prismatic axis and origin, the claim that optimization is unconstrained by VLM limitations survives, and if it does not, the claim collapses.

Watch

Extended reading notes

Core claim

ArtGS's central claim is that the parameters of an articulated object's joints—axis $u_i$, origin $q_i$, and type $c_i$ for each movable part $i$—can be optimized to high accuracy by coupling a static 3D Gaussian Splatting reconstruction with a rigid skeleton $B(\theta)$ whose $SE(3)$ bone transformations are revolute (Rodrigues rotation) or prismatic (translation). The skeleton is initialized by a fine-tuned vision-language model that outputs part bounding boxes and joint types, then refined by minimizing $L = \lambda_{L1} L_1 + \lambda_{SSIM} L_{SSIM} + L_{reg}$ over the joint parameters $J$ and per-frame object states $\{\theta_t\}$, using RGB-D frames recorded while the robot manipulates the object under impedance control. Gaussian means and rotations are deformed by linear blend skinning with learnable per-point weights, which makes the render-and-compare loop differentiable. In the paper's experiments this brings joint axis error below 3 degrees in most categories and lifts simulated manipulation success to 62–90 percent, outperforming the compared point-cloud and interaction-based baselines.

Load-bearing premise

The pipeline assumes the vision-language model identifies the correct number of movable parts and the correct joint type (revolute or prismatic) for every part, since the skeleton $B(\theta)$ represents only those two joint types and a wrong part count or joint type cannot be recovered by the rendered-image loss.

Editorial extensions

If this is right

  • A robot can correct its own initial perception errors: in several categories ArtGS takes VLM joint-axis errors above 20 degrees down to 1–3 degrees after interaction.
  • Manipulation policies driven by the refined joints succeed at 62–90 percent across seven simulated categories, versus 42–57 percent for the strongest compared baseline.
  • Real-world trials on five objects improve from roughly 4–7 successes per 10 tries without optimization to 7–10 per 10 tries with the closed loop.
  • Because each robot link is reconstructed separately with MDH kinematics, the same representation and optimization pipeline transfers to different robotic arms such as Franka and xArm7.
  • The method needs no CAD model or category-level joint prior at test time: only multi-view RGB-D images and a correct joint-type and part-count initialization from the VLM.

Reading between the lines

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

  • The paper's own Table II shows the VLM-only run (ArtGS w/o opt.) often underperforms Ditto before optimization, so the headline gains are attributable to the interactive refinement loop rather than to the VLM's initial guesses.
  • Because the skeleton supports only revolute and prismatic joints, the central claim does not yet cover screw joints, universal joints, or deformable parts; extending it would require a richer bone model and revalidation.
  • A testable extension is to use the same Gaussian-based robot reconstruction to calibrate the robot's own MDH parameters from its motion, turning the manipulation loop into a self-calibration procedure.
  • Since the VLM determines part count and joint type, the practical ceiling of the pipeline is set by the VLM's classification accuracy; a geometric fallback that proposes multiple joint hypotheses would let the optimization itself decide the joint type.
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

4 major / 5 minor

Summary. The paper proposes ArtGS, a pipeline that combines static 3D Gaussian Splatting reconstruction with vision-language-model-based initial joint/bone estimation and a dynamic differentiable-rendering optimization stage. Given multi-view RGB-D observations of an articulated object, ArtGS reconstructs the object as Gaussians, uses a fine-tuned VLM to propose movable parts and joint types/axes, then refines the joint parameters and per-frame object states by minimizing a rendering loss (Eq. 9) against images and depth maps collected during robot manipulation. Experiments are reported in SAPIEN simulation and on a real Franka robot for seven object categories and five real-world objects, with claims of state-of-the-art joint estimation accuracy and manipulation success rates, plus a cross-embodiment demonstration with Franka and xArm7.

Significance. If the empirical claims are correct, ArtGS would be a valuable integration: the idea of using the VLM only as an initializer and relying on differentiable 3DGS rendering for closed-loop refinement of articulation parameters is sound and practically appealing, and the cross-embodiment digital-twin idea is a useful direction for simulation-to-real transfer. The paper also names concrete components—the MDH-based robot rendering, the impedance-control interaction loop, and the rendering-based correction of large VLM axis errors—that could benefit the community. However, the current evidence is not sufficient to support the strong empirical claims. The central robustness claim depends on an untested structural assumption about VLM-provided joint type and part count, the experimental tables are internally inconsistent, and no statistical or trial-count information is given for the main simulation results. The paper ships no code or detailed supplementary material, so the reported numbers cannot be independently checked at this stage.

major comments (4)
  1. [Sec. V-A, Table II] The statement that ArtGS 'consistently outperforms all other methods' is contradicted by the paper's own table. In Table II, Ditto reports lower axis error (AE) than ArtGS for Refrigerator (0.46 vs. 1.10), Storage (0.31 vs. 1.52), and Drawer (5.06 vs. 8.32). The table also has an inconsistent number of entries per row (13 instead of 14), with the OE value for Cabinet missing for every method, and the footnote refers to a '*' marker that does not appear in the table. This undermines the central quantitative claim and must be corrected and re-verified.
  2. [Sec. V-B, Table III] Table III reports success rates as single percentages without any indication of the number of trials, variance, or statistical significance. Consequently, the claim that ArtGS 'significantly outperforms' the baselines cannot be assessed. Even Table IV, which does state 10 trials per object, provides no confidence intervals or pairwise significance tests. Please report trial counts and dispersion measures for all main results, and apply an appropriate significance test for the pairwise comparisons.
  3. [Sec. IV-B/C, Sec. V-C, Eq. (7), Eq. (9)] The pipeline's ability to recover articulation parameters is structurally limited by the VLM's discrete outputs. The skeleton B(θ) in Eq. (7) only generates revolute and prismatic SE(3) transforms, and the skinning weights W∈R^{K+1} fix the number of movable parts K; Eq. (9) optimizes continuous joint parameters and states but not the discrete joint type c_i or K. Therefore, if the VLM mispredicts a joint type or the part count, the true articulation lies outside the representable model class and the rendering loss cannot in principle drive the parameters to the ground truth. The ablation text in Sec. V-C appropriately qualifies the 'not constrained by VLM limitations' claim with 'when the VLM accurately predicts the joint type and axis', but the paper gives no confusion matrix for VLM joint-type or part-count accuracy, and all seven experimental categories are revolute/prismatic. Please report VLM joint-type and part-count accuracy separately, and either restrict the robustness claim accordingly or extend the model to recover discrete structure.
  4. [Sec. V, Datasets and Real-World Experiment] There is a direct factual inconsistency in the description of the experimental setup. Sec. V says the dataset categories are 'storage (revolute joints), drawer (prismatic joints)', while the real-world paragraph states the real objects include 'storage (prismatic joints), drawer (revolute joints)', and Table III labels the storage category 'revolute' and the drawer category 'prismatic'. Please correct this inconsistency and ensure all tables, text, and figure captions use a single consistent labeling of joint types.
minor comments (5)
  1. [Sec. III, Eq. (8)] The notation for skinning weights is inconsistent: the text defines W∈R^{K+1}, but each Gaussian is said to have an attribute W_i, and Eq. (8) uses W_{ij}. Please clarify whether the weights are per-Gaussian vectors or a global matrix, and index them consistently.
  2. [Sec. IV-C, Eq. (8)] Applying linear blend skinning to rotation factors r_i∈SO(3) by a weighted sum of SE(3) matrices is not a standard rotation interpolation and may produce non-orthogonal results. Please specify the exact normalization or interpolation used, or state that r_i is treated as a homogeneous vector in the LBS formula.
  3. [Table I] The task table has formatting problems: the columns are not separated clearly and some cells run together (e.g., 'This is a object categorywithnumber of movable parts'). Please reformat so that each capability, task, and example template is a distinct cell.
  4. [Throughout] There are several typos and spacing errors, including 'P RELIMINARY' in the section heading, 'V AT-Mart' in Sec. V-B and the references, and 'Commitee' in the acknowledgment. Please proofread the manuscript.
  5. [Sec. V-A] The baseline description for Ditto says it was tested 'using a 30° and 30 cm configuration', but the relationship of these values to the reported errors is not explained; please specify what these parameters mean and how they affect the comparison.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation found: ArtGS's joint-parameter optimization is evaluated against external RGB-D observations and measurably corrects the VLM's continuous axis/origin errors, while the VLM-supplied part count and joint type are a stated conditioning assumption rather than a fitted input renamed as a prediction.

full rationale

The central derivation is the dynamic optimization in Sec. IV-C, which minimizes L = lambda_L1 * L1 + lambda_SSIM * LSSIM + Lreg (Eq. 9) over J = {psi_i} and {theta_t} by comparing rendered images and depth maps to the multi-view RGB-D images {I_t} collected during manipulation. This loss is external to the VLM: ArtGS (w/o opt.) shows axis errors above 20 degrees on dishwasher, refrigerator, and drawer, while ArtGS reduces them to 1-3 degrees on most categories (Table II), so the refined parameters are not the VLM's outputs by construction. The only structural quantities inherited from the VLM are the part count K and discrete joint type c_i, since the skeleton B(theta) in Eq. 7 and the skinning weights W in R^{K+1} fix those choices; the paper itself conditions the claim on this, stating in Sec. V-C that 'when the VLM accurately predicts the joint type and axis, our method achieves superior optimization results in the majority of cases.' That is a robustness limitation, not circularity, because the optimization loss is still evaluated against observed images and does not reduce to the VLM's own estimates. Self-citations (GAMMA, RPMArt, ReArtGS, etc.) appear as baselines or related work and are not load-bearing premises; no uniqueness theorem or ansatz is imported from the authors' prior work to force the result. The system is benchmarked against external methods and external observations, so no circular step is established.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on standard 3DGS, LBS, and impedance-control models, plus the assumption that the VLM's structural guesses are correctable by optimization. No new physical entities are postulated.

free parameters (2)
  • Loss weights lambda_L1 and lambda_SSIM = not reported
    These hyperparameters in Eq. 9 balance color and structure losses; their values affect how joint parameters are optimized and are not disclosed.
  • Gaussian count n in static reconstruction = not reported
    The number of Gaussian spheres in the scene is determined by the 3DGS training but ultimately controlled by hyperparameters; reconstruction quality feeds into the VLM initialization and subsequent optimization.
assumptions (4)
  • domain assumption Objects consist of K rigid movable parts connected by revolute or prismatic joints.
    The model in Section IV defines joints only as revolute or prismatic; any joint type outside this set cannot be represented by the skeleton B(theta).
  • domain assumption Linear blend skinning with learnable weights maps canonical Gaussians to deformed poses accurately.
    Eq. 8 assumes each Gaussian follows a weighted combination of rigid joint transformations; this is a standard skinning model and assumes no non-rigid deformation.
  • domain assumption The differentiable renderer provides usable gradients of the image loss with respect to joint parameters J.
    The closed-loop optimization in Section IV-C depends on gradient flow through rendering to update axis and origin parameters; the paper does not include a formal proof or convergence analysis.
  • domain assumption The fine-tuned VLM supplies correct joint type and part count for each object.
    If the VLM misclassifies joint type or number of parts, the skeleton representation in Eq. 7 cannot express the true articulation, and refinement cannot recover from the error.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ArtGS:3D Gaussian Splatting for Interactive Visual-Physical Modeling and Manipulation of Articulated Objects." pith.science (2026). https://pith.science/paper/W3UHNMOW

@misc{pith2026250702600,
  author       = {Pith},
  title        = {Pith review of: ArtGS:3D Gaussian Splatting for Interactive Visual-Physical Modeling and Manipulation of Articulated Objects},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W3UHNMOW}},
  note         = {Machine review of arXiv:2507.02600}
}
read the original abstract

Articulated object manipulation remains a critical challenge in robotics due to the complex kinematic constraints and the limited physical reasoning of existing methods. In this work, we introduce ArtGS, a novel framework that extends 3D Gaussian Splatting (3DGS) by integrating visual-physical modeling for articulated object understanding and interaction. ArtGS begins with multi-view RGB-D reconstruction, followed by reasoning with a vision-language model (VLM) to extract semantic and structural information, particularly the articulated bones. Through dynamic, differentiable 3DGS-based rendering, ArtGS optimizes the parameters of the articulated bones, ensuring physically consistent motion constraints and enhancing the manipulation policy. By leveraging dynamic Gaussian splatting, cross-embodiment adaptability, and closed-loop optimization, ArtGS establishes a new framework for efficient, scalable, and generalizable articulated object modeling and manipulation. Experiments conducted in both simulation and real-world environments demonstrate that ArtGS significantly outperforms previous methods in joint estimation accuracy and manipulation success rates across a variety of articulated objects. Additional images and videos are available on the project website: https://sites.google.com/view/artgs/home

Figures

Figures reproduced from arXiv: 2507.02600 by the authors.

Figure 1
Figure 1. The overview of ArtGS. (a) Multi-view observations of the articulated object. (b) Static reconstruction using 3DGS. (c) Inference of articulated bone parameters using the fine-tuned Vision-Language Model. (d) Optimization of parameters through robotic manipulation. spatial information, the inherently sparse and unordered nature of point clouds—along with their limited temporal consistency—continues to pose significa… view at source ↗
Figure 2
Figure 2. Pipeline of ArtGS. Starting from multi-view RGB-D inputs and object masks, ArtGS performs static 3DGS reconstruction and synthesizes robot poses. It then uses a VLM-based bone initialization module to infer the articulated bone parameters through visual-language reasoning. Finally, the Bone Refinement module dynamically optimizes the parameters of revolute and prismatic parts, producing a precise kinematic model of … view at source ↗
Figure 3
Figure 3. Settings of Real-world Environment. Two cameras are used: an eye-in-hand camera for capturing multi-view images in static 3D GS reconstruction and a fixed third-person camera for single-view images to update the dynamic 3D GS. we can continuously improve the modeling accuracy of the articulated objects. V. EXPERIMENT In this section, we conduct comprehensive experiments on modeling and manipulation in both simulated… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Cross-Embodiment Experiment. This figure demonstrates the cross-embodiment capability of ArtGS. The first and second rows show qualitative results for the Franka and xArm7 robotic arms, respectively. The first column displays the robotic arm reconstruction results from…
Figure 5
Figure 5. Figure 5: Real-world Experiment. We implement ArtGS and only fine-tuned VLM in the real-world experiments. Manipulation tasks include opening the door of cabinet (2 parts), drawer, storage, and microwave challenges during manipulation. Furthermore, ArtGS demon￾strates robust cro…

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

  1. PokeNet: Learning Kinematic Models of Articulated Objects from Human Observations

    cs.RO 2026-02 conditional novelty 6.0 of 10

    PokeNet estimates joint types, axes, ranges, and operation order of articulated objects directly from a single-view point cloud video of a human demonstration.

  2. PhysGraph: A Physics-aware 3D Scene Graph for Perception and Reasoning

    cs.RO 2026-06 unverdicted novelty 5.0 of 10

    PhysGraph reconstructs object-centric 3D geometry from RGB-D, decomposes objects into parts, infers materials and articulations via visual reasoning, and reports SOTA results on semantic segmentation, multi-object mas...

Reference graph

Works this paper leans on

41 extracted references · 27 canonical work pages · cited by 2 Pith papers

  1. [1]

    Vat-mart: Learning visual action trajectory proposals for manipulating 3d articulated objects,

    R. Wu, Y . Zhao, K. Mo, Z. Guo, Y . Wang, T. Wu, Q. Fan, X. Chen, L. Guibas, and H. Dong, “Vat-mart: Learning visual action trajectory proposals for manipulating 3d articulated objects,” arXiv preprint arXiv:2106.14440, 2021

    arXiv 2021

  2. [2]

    Robotwin: Dual-arm robot benchmark with generative digital twins,

    Y . Mu, T. Chen, Z. Chen, S. Peng, Z. Lan, Z. Gao, Z. Liang, Q. Yu, Y . Zou, M. Xu et al. , “Robotwin: Dual-arm robot benchmark with generative digital twins,” in Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 27 649–27 660

    work page 2025

  3. [3]

    Forcevla: Enhancing vla models with a force-aware moe for contact-rich manipulation,

    J. Yu, H. Liu, Q. Yu, J. Ren, C. Hao, H. Ding, G. Huang, G. Huang, Y . Song, P. Cai et al. , “Forcevla: Enhancing vla models with a force-aware moe for contact-rich manipulation,” arXiv preprint arXiv:2505.22159, 2025

    arXiv 2025

  4. [4]

    Gamma: Generalizable articulation modeling and manipulation for articulated objects,

    Q. Yu, J. Wang, W. Liu, C. Hao, L. Liu, L. Shao, W. Wang, and C. Lu, “Gamma: Generalizable articulation modeling and manipulation for articulated objects,” in 2024 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2024, pp. 5419–5426

    work page 2024

  5. [5]

    Rpmart: Towards robust perception and manipulation for articulated objects,

    J. Wang, W. Liu, Q. Yu, Y . You, L. Liu, W. Wang, and C. Lu, “Rpmart: Towards robust perception and manipulation for articulated objects,” in 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2024, pp. 7270–7277

    work page 2024

  6. [6]

    3d gaussian splatting for real-time radiance field rendering

    B. Kerbl, G. Kopanas, T. Leimkühler, and G. Drettakis, “3d gaussian splatting for real-time radiance field rendering.” ACM Trans. Graph., vol. 42, no. 4, pp. 139–1, 2023

    work page 2023

  7. [7]

    Category-Level Articulated Object Pose Estimation,

    X. Li, H. Wang, L. Yi, L. J. Guibas, A. L. Abbott, and S. Song, “Category-Level Articulated Object Pose Estimation,” in 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, pp. 3703–3712

    work page 2020

  8. [8]

    Neural Deformation Graphs for Globally-consistent Non-rigid Recon- struction,

    A. Bozic, P. Palafox, M. Zollhofer, J. Thies, A. Dai, and M. Niesner, “Neural Deformation Graphs for Globally-consistent Non-rigid Recon- struction,” in 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, pp. 1450–1459

    work page 2021

Show all 41 references
  1. [9]

    Reartgs: Reconstructing and generating articulated objects via 3d gaussian splatting with geometric and motion constraints,

    D. Wu, L. Liu, Z. Linli, A. Huang, L. Song, Q. Yu, Q. Wu, and C. Lu, “Reartgs: Reconstructing and generating articulated objects via 3d gaussian splatting with geometric and motion constraints,” arXiv preprint arXiv:2503.06677, 2025

    2025

  2. [10]

    MARS: Multimodal Active Robotic Sensing for Articulated Characterization,

    H. Zeng, P. Zhang, C. Wu, J. Wang, T. Ye, and F. Li, “MARS: Multimodal Active Robotic Sensing for Articulated Characterization,” in Proceedings of the Thirty-ThirdInternational Joint Conference on Artificial Intelligence. International Joint Conferences on Artificial Intellige...

  3. [11]

    Ditto: Building digital twins of articulated objects from interaction,

    Z. Jiang, C.-C. Hsu, and Y . Zhu, “Ditto: Building digital twins of articulated objects from interaction,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2022, pp. 5616–5626

    2022

  4. [12]

    Neural implicit representation for building digital twins of unknown articulated objects,

    Y . Weng, B. Wen, J. Tremblay, V . Blukis, D. Fox, L. J. Guibas, and S. T. Birchfield, “Neural implicit representation for building digital twins of unknown articulated objects,” 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pp. 3141–3150, 2024

    2024

  5. [13]

    Screwnet: Category- independent articulation model estimation from depth images using screw theory,

    A. Jain, R. Lioutikov, C. Chuck, and S. Niekum, “Screwnet: Category- independent articulation model estimation from depth images using screw theory,” in 2021 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2021, pp. 13 670–13 677

    2021

  6. [14]

    Visual identification of articulated object parts,

    V . Zeng, T. E. Lee, J. Liang, and O. Kroemer, “Visual identification of articulated object parts,” in 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . IEEE, 2021, pp. 2443–2450

    2021

  7. [15]

    Uniaff: A unified representation of affordances for tool usage and articulation with vision-language models,

    Q. Yu, S. Huang, X. Yuan, Z. Jiang, C. Hao, X. Li, H. Chang, J. Wang, L. Liu, H. Li et al., “Uniaff: A unified representation of affordances for tool usage and articulation with vision-language models,” arXiv preprint arXiv:2409.20551, 2024

    2024 arXiv

  8. [16]

    Command- driven Articulated Object Understanding and Manipulation,

    R. Chu, Z. Liu, X. Ye, X. Tan, X. Qi, C.-W. Fu, and J. Jia, “Command- driven Articulated Object Understanding and Manipulation,” in 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, pp. 8813–8823

    2023

  9. [17]

    Generalizable articulated object perception with superpoints,

    Q. Yu, C. Hao, X. Yuan, L. Zhang, L. Liu, Y . Huo, R. Agarwal, and C. Lu, “Generalizable articulated object perception with superpoints,” in ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) . IEEE, 2025, pp. 1–5

    2025

  10. [18]

    Gapartnet: Cross-category domain-generalizable object perception and manipulation via generalizable and actionable parts,

    H. Geng, H. Xu, C. Zhao, C. Xu, L. Yi, S. Huang, and H. Wang, “Gapartnet: Cross-category domain-generalizable object perception and manipulation via generalizable and actionable parts,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023,...

    2023

  11. [19]

    Unsupervised Pose-aware Part Decomposition for Man-Made Articulated Objects,

    Y . Kawana, Y . Mukuta, and T. Harada, “Unsupervised Pose-aware Part Decomposition for Man-Made Articulated Objects,” in Computer Vision – ECCV 2022, S. Avidan, G. Brostow, M. Cissé, G. M. Farinella, and T. Hassner, Eds. Springer Nature Switzerland, vol. 13663, pp. 558–575

    2022

  12. [20]

    Neural articulated radiance field,

    A. Noguchi, X. Sun, S. Lin, and T. Harada, “Neural articulated radiance field,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 5762–5772

    2021

  13. [21]

    Learning implicit representation for reconstructing articulated objects,

    H. Zhang, F. Li, S. Rawlekar, and N. Ahuja, “Learning implicit representation for reconstructing articulated objects,” arXiv preprint arXiv:2401.08809, 2024

    2024 arXiv

  14. [22]

    Deformable 3d gaussian splatting for animatable human avatars,

    H. Jung, N. Brasch, J. Song, E. Perez-Pellitero, Y . Zhou, Z. Li, N. Navab, and B. Busam, “Deformable 3d gaussian splatting for animatable human avatars,” arXiv preprint arXiv:2312.15059 , 2023

    2023 arXiv

  15. [23]

    Gart: Gaussian articulated template models,

    J. Lei, Y . Wang, G. Pavlakos, L. Liu, and K. Daniilidis, “Gart: Gaussian articulated template models,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2024, pp. 19 876–19 887

    2024

  16. [24]

    Maniskill2: A unified benchmark for generalizable manipulation skills,

    J. Gu, F. Xiang, X. Li, Z. Ling, X. Liu, T. Mu, Y . Tang, S. Tao, X. Wei, Y . Yao et al. , “Maniskill2: A unified benchmark for generalizable manipulation skills,” arXiv preprint arXiv:2302.04659 , 2023

    2023 arXiv

  17. [25]

    Instruction-driven history-aware policies for robotic manipulations,

    P.-L. Guhur, S. Chen, R. G. Pinel, M. Tapaswi, I. Laptev, and C. Schmid, “Instruction-driven history-aware policies for robotic manipulations,” in Conference on Robot Learning . PMLR, 2023, pp. 175–187

    2023

  18. [26]

    Articulated object manipulation using online axis estimation with sam2-based tracking,

    X. Wang, T. Chen, Q. Yu, T. Xu, Z. Chen, Y . Fu, Z. He, C. Lu, Y . Mu, and P. Luo, “Articulated object manipulation using online axis estimation with sam2-based tracking,” arXiv preprint arXiv:2409.16287, 2024

    2024 arXiv

  19. [27]

    Where2act: From pixels to actions for articulated 3d objects,

    K. Mo, L. J. Guibas, M. Mukadam, A. Gupta, and S. Tulsiani, “Where2act: From pixels to actions for articulated 3d objects,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021, pp. 6813–6823

    2021

  20. [28]

    Universal manipulation policy network for articulated objects,

    Z. Xu, Z. He, and S. Song, “Universal manipulation policy network for articulated objects,” IEEE robotics and automation letters , vol. 7, no. 2, pp. 2447–2454, 2022

    2022

  21. [29]

    Articulated object interaction in unknown scenes with whole-body mobile manipulation,

    M. Mittal, D. Hoeller, F. Farshidian, M. Hutter, and A. Garg, “Articulated object interaction in unknown scenes with whole-body mobile manipulation,” in 2022 IEEE/RSJ international conference on intelligent robots and systems (IROS) . IEEE, 2022, pp. 1647–1654

    2022

  22. [30]

    Nerf: Representing scenes as neural radiance fields for view synthesis,

    B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoor- thi, and R. Ng, “Nerf: Representing scenes as neural radiance fields for view synthesis,” Communications of the ACM , vol. 65, no. 1, pp. 99–106, 2021

    2021

  23. [31]

    Splatsim: Zero-shot sim2real transfer of rgb manipulation policies using gaussian splatting,

    M. N. Qureshi, S. Garg, F. Yandun, D. Held, G. Kantor, and A. Silwal, “Splatsim: Zero-shot sim2real transfer of rgb manipulation policies using gaussian splatting,” arXiv preprint arXiv:2409.10161 , 2024

    2024 arXiv

  24. [32]

    Robo-gs: A physics consistent spatial-temporal model for robotic arm with hybrid representation,

    H. Lou, Y . Liu, Y . Pan, Y . Geng, J. Chen, W. Ma, C. Li, L. Wang, H. Feng, L. Shi et al., “Robo-gs: A physics consistent spatial-temporal model for robotic arm with hybrid representation,” arXiv preprint arXiv:2408.14873, 2024

    2024 arXiv

  25. [33]

    Manigaussian: Dynamic gaussian splatting for multi-task robotic manipulation,

    G. Lu, S. Zhang, Z. Wang, C. Liu, J. Lu, and Y . Tang, “Manigaussian: Dynamic gaussian splatting for multi-task robotic manipulation,” in European Conference on Computer Vision . Springer, 2024, pp. 349– 366

    2024

  26. [34]

    Gaussiangrasper: 3d language gaussian splatting for open-vocabulary robotic grasping,

    Y . Zheng, X. Chen, Y . Zheng, S. Gu, R. Yang, B. Jin, P. Li, C. Zhong, Z. Wang, L. Liu et al., “Gaussiangrasper: 3d language gaussian splatting for open-vocabulary robotic grasping,” IEEE Robotics and Automation Letters, 2024

    2024

  27. [35]

    Differentiable robot rendering,

    R. Liu, A. Canberk, S. Song, and C. V ondrick, “Differentiable robot rendering,” arXiv preprint arXiv:2410.13851 , 2024

    2024 arXiv

  28. [36]

    Robogsim: A real2sim2real robotic gaussian splatting simulator,

    X. Li, J. Li, Z. Zhang, R. Zhang, F. Jia, T. Wang, H. Fan, K.-K. Tseng, and R. Wang, “Robogsim: A real2sim2real robotic gaussian splatting simulator,” arXiv preprint arXiv:2411.11839 , 2024

    2024 arXiv

  29. [37]

    A simple and systematic approach to assigning denavit– hartenberg parameters,

    P. I. Corke, “A simple and systematic approach to assigning denavit– hartenberg parameters,” IEEE transactions on robotics , vol. 23, no. 3, pp. 590–594, 2007

    2007

  30. [38]

    Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks,

    Z. Chen, J. Wu, W. Wang, W. Su, G. Chen, S. Xing, M. Zhong, Q. Zhang, X. Zhu, L. Lu et al., “Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks,” in Proceed- ings of the IEEE/CVF conference on computer vision and pattern recognition,...

    2024

  31. [39]

    Sapien: A simulated part-based interactive en- vironment,

    F. Xiang, Y . Qin, K. Mo, Y . Xia, H. Zhu, F. Liu, M. Liu, H. Jiang, Y . Yuan, H. Wanget al., “Sapien: A simulated part-based interactive en- vironment,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2020, pp. 11 097–11 107

    2020

  32. [40]

    Category-level articulated object pose estimation,

    X. Li, H. Wang, L. Yi, L. J. Guibas, A. L. Abbott, and S. Song, “Category-level articulated object pose estimation,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2020, pp. 3706–3715

    2020

  33. [41]

    Addressing function approxi- mation error in actor-critic methods,

    S. Fujimoto, H. Hoof, and D. Meger, “Addressing function approxi- mation error in actor-critic methods,” in International conference on machine learning. PMLR, 2018, pp. 1587–1596

    2018

Pith tools

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