Pith. sign in

REVIEW 4 major objections 6 minor 28 references

RORA: Realistic Object Reconstruction with Articulation

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read A single static video plus about 42 seconds of user confirmation is enough to reconstruct an articulated, simulation-ready object; the pipeline reads joint axes from boundary geometry and outperforms motion- and vision-language-based…

desk verdict Good pipeline idea, but the evaluation fails to isolate the automatic joint suggestion from the human user, so the headline accuracy numbers are unverifiable. read the letter →

arxiv 2608.04842 v1 pith:HRWKTNLE submitted 2026-08-05 cs.RO cs.GR

classification cs.ROcs.GR
keywords singlestaticvideoarticulatedobjectreconstruction3DGaussiansplattingconvexdecompositionjointaxisestimationhuman-in-the-loopreal-to-simrobotlearning
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

RORA is an attempt to settle a practical question: can an articulated object for physics simulation be reconstructed without filming the object moving? The paper claims yes, through a single static video, a hybrid Gaussian-splatting plus mesh representation, and a short human-in-the-loop step. Convex decomposition splits the mesh into fragments, the user groups fragments into semantic parts in about 42 seconds, and a geometric algorithm proposes joint axes from the boundary curves between parts. The authors report that on a public articulated-object benchmark this beats two baselines, one motion-based and one vision-language-based, on joint accuracy and rendered quality while keeping total runtime around 7.5 minutes. If the claim holds, sim-to-real robot learning gains a cheap route from a single video to a URDF-ready movable asset.

What carries the argument

The load-bearing mechanism is the Automatic Joint Suggestion Algorithm, which turns joint estimation into a local geometry-reading problem. It first finds boundary cycles---closed loops of mesh vertices where one part meets another---by pruning low-degree vertices and walking the remaining graph. For each cycle it fits separating hyperplanes to a local neighborhood: a single linear support-vector machine for planar interfaces, or an expectation-maximization mixture of two planes for corners. The normal vector, the two-plane intersection vector, and their cross product become three candidate axes. Estimates from both sides of an interface are averaged into a bidirectional connection, and heuristic rules add extra candidates when one part pair touches at multiple boundaries, such as the two hinge points of a toolbox handle.

What would settle it

Take a newly collected set of objects whose movable parts are flush with the surrounding surface or very small relative to the object---a recessed drawer, a flush button, a hidden latch---and run RORA as published on a single static video of each. Measure the angle and offset between the user-confirmed joint axes and the ground-truth axes from mechanical drawing measurements; if the suggested axes are consistently wrong or require extensive manual correction on such objects, the central premise that boundary geometry encodes articulation is refuted.

Watch

Extended reading notes

Core claim

The central discovery is that articulation is readable from static geometry: a movable subpart leaves a boundary cycle on the surface, and the local geometry around that cycle encodes the joint axis. The pipeline reconstructs the object's mesh and Gaussian splats from one video, makes the mesh watertight, decomposes it into convex fragments, and asks the user only to group fragments into parts. The Automatic Joint Suggestion Algorithm then extracts boundary cycles, fits one or two separating hyperplanes to the surrounding vertices, and uses the plane normals, the two-plane intersection line, and its cross product as candidate axes. A user selects the correct candidate, and the system exports a URDF with the Gaussians bound to each kinematic link. On the benchmark, this yields near-zero joint type errors, angular errors below about $1.5^\circ$, and position errors under $0.07$ m even in chained and multi-joint categories, with rendering quality above both baselines at unseen configurations.

Load-bearing premise

The method assumes that every movable part sticks out from its neighbour as a geometrically distinguishable bulge, so the interface appears as a closed boundary loop whose surrounding surface shape fixes the joint axis; parts that are flush, recessed, tiny, or hidden inside the object will not be suggested correctly.

Editorial extensions

If this is right

  • A practitioner can produce a simulation-ready articulated asset from a single static video in about 7 minutes 23 seconds, of which only 42 seconds require human attention.
  • Objects with several joints or chained parent-child linkages can be reconstructed accurately, even though their joint trajectories are coupled and would confuse motion-based methods.
  • Because rendering is carried by Gaussians that move with their mesh parts, the asset remains photorealistic when articulated to configurations the scanner never saw.
  • The same pipeline can feed physics simulators directly, enabling real-time dexterous hand teleoperation and robot manipulation data collection without additional asset authoring.

Reading between the lines

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

  • The paper leaves open whether a learned ranker over the suggested candidate axes could replace most of the 42-second human confirmation; the same boundary-cycle features would be a natural input for such a ranker.
  • The stated limitation predicts a sharp failure boundary: flush or tiny subparts. Testing that boundary directly, by scanning objects with recessed drawers or nearly invisible handles, would confirm or refute the geometric premise better than the current benchmark.
  • If the method generalizes, real-to-sim pipelines could adopt a geometry-first division of labour: use static scans and geometric suggestions to build the asset, and use motion capture only to verify or refine the assembled URDF.
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 / 6 minor

Summary. This paper presents RORA, a semi-automated pipeline that turns a single static multi-view video of an object into a simulation-ready articulated asset combining 3D Gaussian Splatting and a mesh representation. The pipeline performs GS2Mesh-based reconstruction and watertight mesh refinement, approximate convex decomposition with user grouping, binding of Gaussians to mesh parts, and an automatic joint suggestion stage that extracts boundary cycles, estimates separating hyperplanes, fuses bidirectional connections, and proposes candidate axes. A GUI lets the user select joint types and make minor adjustments, and the final asset is exported as a URDF with GS-attached links. The authors evaluate on 16 PartNet-Mobility objects across three kinematic-complexity categories against Articulate-Anything and ScrewSplat, reporting lower joint angle/position errors and better rendering metrics, and demonstrate deployment in Unreal Engine and Isaac Sim for dexterous manipulation. The paper states that code and assets will be open-sourced.

Significance. If the quantitative results are taken at face value, the contribution is meaningful: RORA avoids multi-state motion capture for articulation, uses geometry-based joint suggestions that are self-contained rather than fitted to the test set, couples photorealistic rendering with physics-ready meshes, and reports a short 42-second active user time. The weaknesses are primarily in the evaluation: the human-in-the-loop protocol is not specified, the benchmark is small and lacks variance reporting, and the visual comparison protocol is underspecified. These do not invalidate the core idea, but they currently make the headline claims unverifiable from the manuscript.

major comments (4)
  1. [Section IV-A, Tables I and III] The experimental protocol for the human-in-the-loop component is not described. The paper does not state how many users participated, what instructions they received, whether ground-truth joint axes were visible in the GUI, whether users were asked to accept or correct suggestions, or how much the suggested axes were adjusted. Since the reported errors (e.g., 0.220 degrees Category I angle error, 0.849 degrees Category III angle error) are far below what unassisted geometric boundary estimation would be expected to achieve on chained linkages, the numbers in Table I may reflect the user's ability to align to ground truth rather than the quality of the automatic joint suggestion algorithm. This is load-bearing for the claim that RORA 'reconstructs simulation-ready assets with accurate articulation': the paper must either isolate the automatic-suggestion accuracy (e.g., report errors before user adjustment) or provide a full protocol with quantitative user-adjustment statistics. Without this, the comparison against fully automatic baselines is not interpretable.
  2. [Section IV-A, Dataset] The benchmark uses only 16 objects total (6/6/4 per category) and reports category-averaged errors with no standard deviations, per-object breakdown, or confidence intervals. With this sample size, the near-perfect averages in Table I could be driven by a few easy objects, and the 'catastrophic failures' attributed to baselines are not statistically characterized. I request per-object results, variance measures, and a statement of whether the same object-level difficulty labels are available for all methods.
  3. [Section IV-A, Table II] The visual metric comparison is underspecified. It is not stated how the 'reference GT rendering' was generated, which renderer was used for each method, whether identical camera trajectories and lighting were used, or how the test configurations were chosen. Because ScrewSplat produces its own GS representation and Articulate-Anything retrieves meshes, the PSNR/SSIM/LPIPS differences may partly reflect representation and rendering choices rather than articulation accuracy. Please specify the rendering protocol and, if possible, use a common rendering backbone for all methods.
  4. [Section III-C2] The automatic joint suggestion depends on several unspecified hyperparameters (the k-hop neighborhood size, uniqueness threshold tau, number of planes K=2, visibility threshold, KNN k) and on the geometric assumption of extruded parts. Since the paper acknowledges in Section V that non-extruded or very small parts are a failure mode, the evaluation should include a sensitivity analysis for at least tau and k, and should characterize how many benchmark objects fall inside the assumed geometry regime. This would tell the reader whether the reported accuracy generalizes beyond the chosen objects.
minor comments (6)
  1. [Section V] The text 'end-toe-nd' should read 'end-to-end'.
  2. [Introduction] The phrase 'We proposeRealistic Object Reconstruction' is missing a space after 'propose'.
  3. [Related Works] The phrase 'synthesize part shpaes' should read 'synthesize part shapes'.
  4. [Section IV-A] The abstract says 'single static object video input' but the dataset description says 48 multi-view images of a single static configuration; please clarify whether the video is an orbit around the static object and how the 48 frames are sampled.
  5. [Table III] Time notation is inconsistent (05:22 vs. 5 min 22 s; '07:23' lacks a separator pattern consistent with the other rows); please unify the format and report standard deviations across runs.
  6. [References] References [9] and [20] lack venue information, and the citation formatting for [3], [4], and [12] should be made consistent with the rest of the bibliography.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: joint axes are computed from local boundary geometry and benchmarked against external data.

full rationale

RORA's articulation derivation is self-contained. The Automatic Joint Suggestion Algorithm (Section III-C) extracts boundary cycles from the mesh graph, fits separating hyperplanes via SVM or two-plane EM, fuses paired unidirectional connections by averaging, and adds heuristic axes for multi-boundary pairs. These computations do not take ground-truth joint parameters as inputs, and the comparison in Tables I-III is against external PartNet-Mobility objects and two published baselines. No equation-level reduction of an output to an input was found. Self-citations [3], [4], [27], [28] are used for background motivation and simulation deployment, not to justify the joint-derivation claim, so they are not load-bearing. The one legitimate concern is the human-in-the-loop evaluation protocol: Section IV-A and the GUI description in Section III-C4 do not state whether the user had access to ground-truth annotations, so the 0.849-degree and 0.028-m Category III numbers may include human correction rather than automatic estimation. That is an experimental-control gap, not a circular derivation; the paper explicitly claims a suggestion-based human-in-the-loop pipeline rather than a fully automatic predictor. Section V's acknowledged limitations, including non-extruded or tiny subparts and reflective or transparent surfaces, are scope limits, not circular steps.

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

The pipeline introduces no new physical or conceptual entities such as particles or forces. It relies on several domain assumptions about geometry and on a small set of hand-chosen hyperparameters, none of which are fitted to the evaluation data. The most load-bearing assumption is the geometric extrudedness of movable parts.

free parameters (4)
  • visibility_threshold = not reported, configurable
    Ratio used to filter Gaussians as part of the object; hand-chosen and not fitted to the benchmark.
  • uniqueness_threshold tau = 0.5 default
    Greedy cycle selection threshold; hand-set and not learned from data.
  • k in KNN voting = not reported
    Number of nearest mesh vertices used to bind Gaussian points to part IDs; a free choice.
  • number of planes K in EM = 2
    Fixed to two hyperplanes for corner geometry; no justification or ablation for other K.
assumptions (4)
  • domain assumption Movable subparts appear geometrically extruded from the object surface.
    Section III-B: the decomposition stage is 'based on assumption that the movable subparts appear geometrically extruded.' If a part is flush or non-extruded, ACD will not isolate it.
  • domain assumption Joint axes can be estimated from local boundary geometry using separating hyperplanes and their intersections.
    Section III-C: the Automatic Joint Suggestion Algorithm derives candidate axes (normal, intersection, cross product) from local boundary geometry. This fails for mechanisms whose joint axis is not aligned with any boundary-derived direction.
  • domain assumption The reconstructed mesh can be made watertight and geometrically accurate enough for graph analysis.
    Section III-A: watertightness is stated as a prerequisite for graph analysis; this depends on the initial GS2Mesh output and Poisson reconstruction, which can fail on reflective or transparent surfaces as acknowledged in Section V.
  • domain assumption User group assignments correspond to kinematic parts.
    Section III-B: the human assigns group numbers to convex fragments; the accuracy of the final articulation depends on this segmentation being semantically correct.

how reviews work

0 comments
Cite this review

Pith. "Pith review of RORA: Realistic Object Reconstruction with Articulation." pith.science (2026). https://pith.science/paper/HRWKTNLE

@misc{pith2026260804842,
  author       = {Pith},
  title        = {Pith review of: RORA: Realistic Object Reconstruction with Articulation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HRWKTNLE}},
  note         = {Machine review of arXiv:2608.04842}
}
read the original abstract

Replicating real-world environments into simulation by realistic visual representation like NeRF and 3D Gaussian Splatting (3DGS) has emerged as an effective strategy to reduce the sim-to-real gap in robot learning. However, implementing object articulation during the real-to-sim process is still a challenging task. Existing motion tracking or learning based articulation methods shows low success rates on complex kinematic structures having multiple joints. Furthermore, those methods require scan of dynamic motion of objects, which makes reconstruction process much complicated. In this work, we propose the first end-to-end pipeline that reconstructs simulation-ready assets with accurate articulation from a single static object video input through suggestion based human-in-the-loop process. Our approach exports a hybrid representation combining 3DGS for photorealistic rendering and mesh-based geometry for physical interaction. In the reconstruction process, our pipeline performs convex decomposition followed by user grouping for intuitive part segmentation, subsequently binding 3D Gaussians to the corresponding mesh parts. An Automatic Joint Suggestion Algorithm then calculates candidate joint axes from local boundary geometries and presents them to users for efficient articulated asset reconstruction. We have shown that our method achieves precise articulation results on partnet-mobility-v0 dataset and real objects. Additionally we presented a potential usage of our framework on robot learning, deploying the reconstructed assets in Unreal Engine and NVIDIA Isaac Sim, demonstrating real-time dexterous hand manipulation tasks.

Figures

Figures reproduced from arXiv: 2608.04842 by the authors.

Figure 1
Figure 1. The Conceptual overview of RORA framework. Our pipeline recon [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overall pipeline of our reconstruction framework. The method is composed of 6 stages. The process is semi-automated human-in-the-loop process, [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Overview of the Automatic Joint Suggestion Algorithm in Articulation Stage. From the segmented mesh, boundary cycles are extracted. Considering [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: An example of additional candidate generation in the Joint Suggestion [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Additionally, Table III shows the operational runtime [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 5
Figure 5. Figure 5: The rendering comparison on the PartNet-Mobility benchmark. Two representative examples are displayed for each category. For each baseline and our [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Qualitative comparison on real-world objects. Columns depict object motions following joint constraints, where yellow arrows indicate revolute joint [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: The dexterous manipulation demo under the simulation. Upper two [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

28 extracted references · 26 canonical work pages

  1. [1]

    Reconciling reality through simulation: A real-to-sim-to- real approach for robust manipulation,

    M. Torne, A. Simeonov, Z. Li, A. Chan, T. Chen, A. Gupta, and P. Agrawal, “Reconciling reality through simulation: A real-to-sim-to- real approach for robust manipulation,”Robotics: Science and Systems, 2024

  2. [2]

    Re 3sim: Generating high-fidelity simulation data via 3d-photorealistic real-to-sim for robotic manipulation,

    X. Han, J. Yu, M. Liu, Y . Chen, X. Lyu, Y . Tian, B. Wang, W. Zhang, and J. Pang, “Re 3sim: Generating high-fidelity simulation data via 3d-photorealistic real-to-sim for robotic manipulation,” inIn Proc. of International Conference on Robotics and Automation, 2026

  3. [3]

    Human-in-the-loop gaussian splatting for robotic teleoperation,

    Y . Lee, H. Kim, H. Ji, J. Heo, Y . Lee, J. Kang, J. Lee, and D. Lee, “Human-in-the-loop gaussian splatting for robotic teleoperation,”IEEE Robotics and Automation Letters, vol. 11, no. 1, pp. 105–112, 2026

  4. [4]

    Human- in-the-loop object segmentation for 3d gaussian splatting via finger- based vr interface,

    Y . Lee, H. Park, H. Kim, H. Ji, D. Yee, and D. Lee, “Human- in-the-loop object segmentation for 3d gaussian splatting via finger- based vr interface,”Advanced Intelligent Systems, vol. n/a, no. n/a, p. e202501025, 2026

  5. [5]

    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,” inIn Proc. of International Conference on Robotics and Automation, 2025

  6. [6]

    Rl- gsbridge: 3d gaussian splatting based real2sim2real method for robotic manipulation learning,

    Y . Wu, L. Pan, W. Wu, G. Wang, Y . Miao, F. Xu, and H. Wang, “Rl- gsbridge: 3d gaussian splatting based real2sim2real method for robotic manipulation learning,” inIn Proc. of International Conference on Robotics and Automation, 2025, pp. 192–198

  7. [7]

    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,” 2025

  8. [8]

    Articulate-anything: Automatic modeling of articulated objects via a vision-language foundation model,

    L. Le, J. Xie, W. Liang, H.-J. Wang, Y . Yang, Y . J. Ma, K. Vedder, A. Kr- ishna, D. Jayaraman, and E. Eaton, “Articulate-anything: Automatic modeling of articulated objects via a vision-language foundation model,” inIn Proc. of International Conference on Learning Representations, 2025

Show all 28 references
  1. [9]

    Articulate anymesh: Open-vocabulary 3d articulated objects modeling,

    X. Qiu, J. Yang, Y . Wang, Z. Chen, Y . Wang, T.-H. Wang, Z. Xian, and C. Gan, “Articulate anymesh: Open-vocabulary 3d articulated objects modeling,” 2025

  2. [10]

    Urdformer: A pipeline for constructing articulated simulation environments from real-world images,

    Z. Chen, A. Walsman, M. Memmel, K. Mo, A. Fang, K. Vemuri, A. Wu, D. Fox, and A. Gupta, “Urdformer: A pipeline for constructing articulated simulation environments from real-world images,” inIn Proc. of Robotics: Science and Systems, 2024

  3. [11]

    PARIS: Part-level recon- struction and motion analysis for articulated objects,

    J. Liu, A. Mahdavi-Amiri, and M. Savva, “PARIS: Part-level recon- struction and motion analysis for articulated objects,” inIn Proc. of International Conference on Computer Vision, 2023

  4. [12]

    Screwsplat: An end-to-end method for articulated object recognition,

    S. Kim, J. Ha, Y . H. Kim, Y . Lee, and F. C. Park, “Screwsplat: An end-to-end method for articulated object recognition,” inIn Proc. of Conference on Robot Learning, 2025

  5. [13]

    Neural signed distance function inference through splatting 3d gaussians pulled on zero-level set,

    W. Zhang, Y .-S. Liu, and Z. Han, “Neural signed distance function inference through splatting 3d gaussians pulled on zero-level set,” in In Proc. of Advances in Neural Information Processing Systems, 2024

  6. [14]

    Sparse voxels rasterization: Real-time high-fidelity radiance field rendering,

    C. Sun, J. Choe, C. Loop, W.-C. Ma, and Y .-C. F. Wang, “Sparse voxels rasterization: Real-time high-fidelity radiance field rendering,” inIn Proc. of Conference on Computer Vision and Pattern Recognition, vol. abs/2412.04459, 2025

  7. [15]

    Gsdf: 3dgs meets sdf for improved rendering and reconstruction,

    M. Yu, T. Lu, L. Xu, L. Jiang, Y . Xiangli, and B. Dai, “Gsdf: 3dgs meets sdf for improved rendering and reconstruction,” inIn Proc. of Advances in Neural Information Processing Systems, 2024

  8. [16]

    GS2Mesh: Surface reconstruction from Gaussian splatting via novel stereo views,

    Y . Wolf, A. Bracha, and R. Kimmel, “GS2Mesh: Surface reconstruction from Gaussian splatting via novel stereo views,” inIn Proc. of European Conference on Computer Vision, 2024

  9. [17]

    Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering,

    A. Gu ´edon and V . Lepetit, “Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering,” inIn Proc. of Conference on Computer Vision and Pattern Recognition, 2024, pp. 5354–5363

  10. [18]

    Vr-gs: A physical dynamics-aware interactive gaussian splatting system in virtual reality,

    Y . Jiang, C. Yu, T. Xie, X. Li, Y . Feng, H. Wang, M. Li, H. Lau, F. Gao, Y . Yang, and C. Jiang, “Vr-gs: A physical dynamics-aware interactive gaussian splatting system in virtual reality,”In Proc. of SIGGRAPH Conference, 2024

  11. [19]

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

    Y . Weng, B. Wen, J. Tremblay, V . Blukis, D. Fox, L. Guibas, and S. Birchfield, “Neural implicit representation for building digital twins of unknown articulated objects,” 2024

  12. [20]

    Urdf-anything+: Autoregressive articulated 3d models generation for physical simulation,

    Z. Wu, Y . Xin, C. Hou, M. Chen, Y . Lyu, J. Zhang, and S. Zhang, “Urdf-anything+: Autoregressive articulated 3d models generation for physical simulation,” 2026

  13. [21]

    Marching cubes: A high resolution 3d surface construction algorithm,

    W. E. Lorensen and H. E. Cline, “Marching cubes: A high resolution 3d surface construction algorithm,” inIn Proc. of Conference on Computer Graphics and Interactive Techniques, ser. SIGGRAPH ’87. New York, NY , USA: Association for Computing Machinery, 1987, p. 163–169

  14. [22]

    Sam 2: Segment anything in images and videos,

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

  15. [23]

    Screened poisson surface reconstruction,

    M. Kazhdan and H. Hoppe, “Screened poisson surface reconstruction,” vol. 32, no. 3, Jul. 2013

  16. [24]

    Garland and P

    M. Garland and P. S. Heckbert,Surface Simplification Using Quadric Error Metrics, 1st ed. New York, NY , USA: Association for Computing Machinery, 1997

  17. [25]

    Approximate convex decomposition for 3d meshes with collision-aware concavity and tree search,

    X. Wei, M. Liu, Z. Ling, and H. Su, “Approximate convex decomposition for 3d meshes with collision-aware concavity and tree search,”ACM Trans. Graph., vol. 41, no. 4, Jul. 2022

  18. [26]

    SAPIEN: A simulated part-based interactive environment,

    F. Xiang, Y . Qin, K. Mo, Y . Xia, H. Zhu, F. Liu, M. Liu, H. Jiang, Y . Yuan, H. Wang, L. Yi, A. X. Chang, L. J. Guibas, and H. Su, “SAPIEN: A simulated part-based interactive environment,” inIn Proc. of Conference on Computer Vision and Pattern Recognition, June 2020

  19. [27]

    Gpu-accelerated subsystem-based admm for large-scale inter- active simulation,

    H. Ji, H. Kim, J. Lee, S. Lee, S. An, J. Heo, Y . Lee, Y . Lee, and D. Lee, “Gpu-accelerated subsystem-based admm for large-scale inter- active simulation,” in2025 IEEE International Conference on Robotics and Automation (ICRA), 2025, pp. 13 868–13 874

  20. [28]

    Visual-inertial hand motion tracking with robustness against occlusion, interference, and contact,

    Y . Lee, W. Do, H. Yoon, J. Heo, W. Lee, and D. Lee, “Visual-inertial hand motion tracking with robustness against occlusion, interference, and contact,”Science Robotics, vol. 6, no. 58, p. eabe1315, 2021

Pith tools

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