PaperlayerView exact revision on arXiv

Source-linked AI summary

Particulate: Feed-Forward 3D Object Articulation

Ruining Li, Yuxin Yao, Chuanxia Zheng, Christian Rupprecht, Joan Lasenby, Shangzhe Wu, Andrea Vedaldi

arXiv:2512.11798v2cs.CVcs.AIcs.GR

TL;DR

Estimating articulated structure from a static 3D mesh remains challenging across the variety of real-world objects. PARTICULATE predicts parts, kinematics, and motion constraints from a mesh in a feed-forward model, and the authors report faster, more accurate recovery than prior methods. It also handles AI-generated objects, with limitations on unusual articulation configurations and inputs lacking realistic internal structure.

  • Problem

    Estimating an object's articulated structure from a static 3D mesh is challenging because procedural approaches are difficult to scale to the long tail of real-world objects.

    Source: §1 · p. 1

  • Method

    PARTICULATE uses a feed-forward transformer trained on diverse articulated 3D assets to predict mesh parts, kinematic structure, and motion constraints.

    Source: §1 · p. 2, §3.1 · p. 3

  • Results

    PARTICULATE is significantly faster and more accurate than prior methods and generalizes to AI-generated objects.

    Source: §4.2.1 · p. 6, §5 · p. 8

  • Takeaways & Limitations

    The predicted articulated assets can be exported to physics simulators, and the approach can support articulated-object creation from images or text prompts when paired with 3D generation.

    Source: §1 · p. 2, §4.2.1 · p. 6

    PARTICULATE can struggle with articulation configurations unlike its training data and with AI-generated assets lacking realistic internal structures.

    Source: §6.3 · p. 13

Abstract

from arXiv · show

We introduce Particulate, a feed-forward model that, given a 3D mesh of an object, infers its articulations, including its 3D parts, their kinematic structure, and the motion constraints. The model is based on a transformer network, the Part Articulation Transformer, which predicts all these parameters for all joints. We train the network end-to-end on a diverse collection of articulated 3D assets from public datasets. During inference, Particulate maps the output of the network back to the input mesh, yielding a fully articulated 3D model in seconds, much faster than prior approaches that require per-object optimization. Particulate also works on AI-generated 3D assets, enabling the generation of articulated 3D objects from a single (real or synthetic) image when combined with an off-the-shelf image-to-3D model. We further introduce a new challenging benchmark for 3D articulation estimation curated from high-quality public 3D assets, and redesign the evaluation protocol to be more consistent with human preferences. Empirically, Particulate significantly outperforms state-of-the-art approaches.

1. Introduction

Articulated motion is central to understanding many objects and supports uses such as robot manipulation and interactive digital twins. PARTICULATE predicts an object's parts, kinematic structure, and motion constraints from a mesh in a feed-forward pass, alongside a new benchmark and evaluation protocol.

1. Introduction

  • Articulated structure estimation from a single mesh targets objects whose parts move or deform, supporting robot manipulation and interactive digital twins.The paper contrasts this goal with procedural approaches that are difficult to scale to the long tail of real-world objects.Source: §1 · p. 1
  • Prior learning-based methods either segment static objects without modeling articulation or generate articulated objects with limited categories and assumed kinematic structure.Some generation methods also rely on part retrieval to assemble assets.Source: §1 · pp. 1–2
  • PARTICULATE predicts articulated part segmentation, kinematic structure, and motion parameters from a 3D mesh in a single forward pass.The model takes a point cloud approximating the mesh and uses a Part Articulation Transformer trained on diverse public 3D assets.Source: §1 · p. 2
  • The paper introduces a challenging benchmark of 243 high-quality annotated 3D assets and an evaluation protocol with metrics intended to reflect prediction quality more faithfully.The benchmark assets were crafted by Lightwheel under the CC-BY-NC license.Source: §1 · p. 2
  • PARTICULATE's inferred structure can be exported to physics simulators, and the paper reports significant gains using metrics consistent with human preferences.The supplied passage does not provide specific comparison values.Source: §1 · p. 2

2. Related Work

Prior work estimates articulated structure through multi-view optimization, articulated-object generation, or foundation-model-based methods, each with constraints. PARTICULATE instead infers articulation attributes in one feed-forward pass, in seconds.

2. Related Work

  • Lift-from-2D part segmentation is limited by rendered-view coverage and can miss occluded or internal parts, while native 3D models mostly predict semantic parts that may not suit articulation.PARTICULATE jointly predicts articulated parts, kinematic structure, and motion constraints.Source: §2 · p. 2
  • Per-instance NeRF or Gaussian-splatting reconstruction captures object-specific articulation but is slow and requires densely sampled, posed images.These images can be difficult to acquire in practice.Source: §2 · p. 3
  • Articulated-object generative models are often trained on few categories, assume attributes such as kinematic structure, and may rely on part retrieval.Source: §2 · p. 3
  • Foundation-model approaches can generalize broadly across categories, but may be slow per object, struggle with internal or subtle parts, and provide limited multi-joint support.The related-work discussion also distinguishes character-focused automatic rigging from the partly-rigid everyday objects studied here.Source: §2 · p. 3
  • PARTICULATE contrasts with prior approaches by inferring all articulation attributes in one feed-forward pass, recovering objects in seconds rather than hours.The paper reports superior performance across diverse categories.Source: §2 · p. 3

3. Method

PARTICULATE maps a static mesh to articulated parts, their kinematic tree, and joint-motion constraints. A transformer with specialized prediction heads is trained end-to-end, then its predictions are mapped back to the mesh at inference.

3.1. Problem Definition

  • PARTICULATE represents an articulated mesh with part count and segmentation, a rooted kinematic tree, and per-part motion types, directions, axes, and ranges.The tree assumes one base part; motion types include fixed, prismatic, revolute, or both.Source: §3.1 · p. 3

3.2. Part Articulation Transformer

  • A point-cloud transformer combines point tokens with learnable part queries, using alternating query self-attention and query-to-point cross-attention.Point inputs include positions, surface normals, and PartField features; the attention design avoids point-to-point self-attention to support dense clouds with lower memory use.Source: §3.2 · pp. 3–4

3.3. Decoder Heads

  • Specialized heads predict point-to-part assignments, parent-child probabilities, motion types and ranges, prismatic directions, and revolute-axis parameters.Revolute-axis direction is predicted from a part token, while points assigned to that part vote for the axis location.Source: §3.3 · pp. 4–5

3.4. Training

  • Training matches ground-truth parts to distinct queries with Hungarian assignment, then applies a multi-task loss to segmentation, tree structure, and motion parameters.The loss combines segmentation and kinematic-tree terms with motion-constraint losses.Source: §3.4 · p. 5

3.5. Inference

  • At inference, point predictions are assigned to mesh faces by majority vote, active parts are retained, and Edmonds’ algorithm extracts their maximum-spanning arborescence.Motion attributes are read from the active queries; revolute-axis locations use the coordinate-wise median of point votes, and connected-component meshes can receive a segmentation refinement.Source: §3.5 · p. 5

4. Experiments

The experiments evaluate articulated-part segmentation and motion prediction separately, then examine training-data and network-design choices. They also include qualitative results on generated meshes.

4. Experiments

  • The evaluation separates segmentation and motion-prediction comparisons, then uses ablations to assess data and network-design choices.This structure enables comparison with methods that address only one of the two subtasks.Source: §4 · pp. 5–6

4.1. Experiment Details

  • The experiments train on 3,800 objects across 50 categories and evaluate on PartNet-Mobility and a 243-object Lightwheel benchmark.Lightwheel spans 14 categories; the training set discards objects with more than 16 articulated parts.Source: §4.1 · p. 6

4.2.1. Qualitative Results on Synthesized 3D Meshes

  • Qualitative results show predicted articulated parts and motion constraints for generated input meshes, with each object shown in two articulated states.The meshes are generated using an off-the-shelf 3D generator.Source: Figure 3 · p. 7

4.2.2. Comparisons on Articulated Part Segmentation

  • PARTICULATE consistently outperforms all segmentation baselines on both datasets, although all methods except P3SAM drop on the more diverse Lightwheel benchmark.PartField and P3SAM receive the exact ground-truth part count at test time, avoiding unmatched-part penalties.Source: §4.2.2 · p. 7, Table 2 · p. 7
  • For Lightwheel comparisons, semantic-segmentation outputs often miss articulated-part boundaries, while SINGAPO and Articulate AnyMesh have distinct generalization or hidden-part shortcomings.PARTICULATE recovers articulated parts, including internal parts, across a wide variety of object categories.Source: Figure 4 · p. 7, §4.2.2 · p. 7

4.2.3. Comparisons on Articulated Motion Prediction

  • PARTICULATE significantly outperforms all motion-prediction baselines across all metrics on both datasets, including a baseline given ground-truth motion ranges.Evaluation compares fully articulated geometries with every part moved to its maximum extent.Source: §4.2.3 · p. 8, Table 3 · p. 8

4.3. Ablations

  • Reducing training to PartNet-Mobility slightly worsens performance, while revolute-axis over-parameterization and PartField features improve it.The axis design mitigates overfitting and averages errors in individual location estimates; PartField features provide global shape context and richer semantic cues.Source: §4.3 · p. 8, Table 4 · p. 8

5. Conclusion

PARTICULATE directly predicts an articulated structure from a static 3D mesh and is faster and more accurate than prior methods. Its performance is bounded by differences from training data and by imperfections in generated articulated assets.

  • PARTICULATE predicts an object's articulated structure from one static 3D mesh, recovering its parts, kinematic tree, and motion constraints.The feed-forward approach is significantly faster and more accurate than prior methods.Source: §5 · p. 8
  • PARTICULATE works on AI-generated objects and can produce physics-simulator-compatible articulated assets directly from images or text prompts.Source: §5 · p. 8
  • The model struggles with objects whose kinematic structures differ substantially from training examples, which remain far fewer than datasets for images or static 3D meshes.Source: §5 · p. 8
  • Generated articulated assets can have inter-part penetrations from generated-mesh artifacts and imperfect motion predictions.Improving their physical plausibility is identified as future work.Source: §5 · p. 8

6. Additional Experimental Results

Additional evaluations show that PARTICULATE performs well across most benchmark categories, while out-of-distribution categories remain challenging. The section also explains metric choices and identifies failure cases tied to unusual articulations and missing internal structure.

6.1. Additional Details of Evaluation Metrics

  • Without penalties for unmatched parts, the Naive Baseline outperforms all methods on part-wise metrics, motivating penalties for missed small parts.The unpenalized protocols are reported for rest-state and fully articulated geometries.Source: §6.1 · p. 12, Table 5 · p. 12, Table 6 · p. 12

6.2. Per-Category Evaluation Results

  • Articulate AnyMesh handles the unseen stand mixer well, while SINGAPO's oversized predictions inflate range-hood IoU when the base dominates volume.The stand mixer articulation is distinct from the other training objects; range hoods often have a very large base component.Source: §6.2 · pp. 12–13, Table 7 · p. 13

6.3. Failure Cases

  • PARTICULATE can fail on atypical articulation configurations and AI-generated shapes lacking internal structures.The latter creates a distribution shift because training objects contain well-defined internal parts.Source: Figure 5 · p. 13, §6.3 · p. 13

6.2. Per-Category Evaluation Results

  • PARTICULATE outperforms baselines on almost all categories, but performance drops on out-of-distribution stand mixer and stovetop examples.Table 8 reports rest-state and fully articulated-state results by category.Source: §6.2 · p. 12, Table 7 · p. 13, Table 8 · p. 14
Loading 2512.11798v2…