Articulated objects are central to interactive 3D applications, including embodied AI, robotics, and VR/AR, where functional part decomposition and kinematic motion are essential. Yet producing high-fidelity articulated assets remains difficult to scale because it requires reliable part decomposition and kinematic rigging. Existing approaches largely fall into two paradigms: optimization-based reconstruction or distillation, which can be accurate but often takes tens of minutes to hours per instance, and inference-time methods that rely on template or part retrieval, producing plausible results that may not match the specific structure and appearance in the input observation. We introduce a part-centric generative framework for articulated object creation that synthesizes part geometry, composition, and articulation under explicit part-aware conditioning. Our representation models an object as a set of movable parts, each encoded by latent tokens augmented with part identity and articulation cues. Conditioned on a single image, the model generates articulated 3D assets that preserve instance-level correspondence while maintaining valid part structure and motion. The resulting approach avoids per-instance optimization, enables fast feed-forward inference, and supports controllable assembly and articulation, which are important for embodied interaction. Experiments on common articulated categories (e.g., drawers and doors) show improved input consistency, part accuracy, and articulation plausibility over optimization-based and retrieval-driven baselines, while substantially reducing inference time.
Figures & tables
Figure 1. Given a single image, PAct generates an articulated 3D object by predicting a part-decomposed structure, synthesizing high-fidelity part geometry and appearance, and estimating articulation parameters for physics-based simulation. figure description
Figure 2. Pipeline of PAct . Stage 1 predicts a part-decomposed sparse structure from a single image using a Part-Aware Flow Model. Stage 2 refines it into detailed 3D part representations via sparse-transformer denoising, while an articulation module aggregates multi-step features to estimate joint parameters for each part. The predicted joint parameters together with the reconstructed part geometries form the final articulated object.
Figure 3. Architecture of the Stage 1 model. To accommodate multi-part inputs, a subset of attention layers in the original TRELLIS network is repurposed to perform within-part local attention. With this modification, the model is fine-tuned on the target dataset.
Figure 4. Illustration of the Stage 2 pipeline. We cache a subset of intermediate features during part-token denoising. For each part, these features are averaged over selected inference timesteps and then pooled over the token dimension. A learned MLP then predicts the articulation parameters.
Methods \ Metric
RS- dgIoU↓
AS- dgIoU↓
RS- dcDist↓
AS- dcDist↓
RS- dCD↓
AS- dCD↓
AOR ↓
CLIP ↑
SINGAPO ( Liu et al., 2024a )
0.7272
0.7306
0.1585
0.1944
0.0090
0.0256
0.0115
0.8376
ArtFormer ( Su et al., 2025 )
1.3489
1.3876
0.3456
0.4414
0.0396
0.0825
0.0147
0.7690
PhysX3D ( Cao et al., 2025a )
0.6449
0.6464
0.2591
0.3396
0.0209
0.0677
0.0960
0.8132
PhysXAnything ( Cao et al., 2025b )
1.1673
1.1709
0.2965
0.3950
0.0162
0.0481
0.0099
0.8456
Arti-Anything ( Le et al., 2025 )
1.2677
1.2693
0.5632
0.5781
0.2583
0.2803
0.1236
0.8648
Ours + SAM+VLM Mask
0.5316
0.5412
0.1249
0.1656
0.0065
0.0218
0.0248
0.8901
Table 1. Comparison with baselines on the ACD Dataset ( Iliash et al., 2024 ) .
Methods \ Metric
RS- dgIoU↓
AS- dgIoU↓
RS- dcDist↓
AS- dcDist↓
RS- dCD↓
AS- dCD↓
AOR ↓
CLIP ↑
SINGAPO ( Liu et al., 2024a )
0.5119
0.5166
0.1177
0.1505
0.0040
0.0202
0.0097
0.8779
ArtFormer ( Su et al., 2025 )
1.3165
1.3212
0.3542
0.5099
0.0504
0.1849
0.0191
0.7877
PhysX3D ( Cao et al., 2025a )
0.7610
0.7637
0.3233
0.4562
0.0115
0.1204
0.2125
0.7972
PhysXAnything ( Cao et al., 2025b )
0.9596
0.9661
0.2538
0.4034
0.0097
0.1106
0.0056
0.8859
Arti-Anything ( Le et al., 2025 )
0.6865
0.6969
0.1386
0.3414
0.0107
0.1230
0.0056
0.8795
Ours + SAM+VLM Mask
0.2485
0.2736
0.0707
0.1191
0.0019
0.0176
0.0233
0.8996
Table 2. Comparison with baselines on the PartNet-Mobility Dataset ( Xiang et al., 2020 ) .
Figure 5. Qualitative comparison on the PartNet-Mobility ( Xiang et al., 2020 ) and ACD ( Iliash et al., 2024 ) datasets. We compare PAct with SINGAPO ( Liu et al., 2024a ) , Articulate-Anything ( Le et al., 2025 ) .
Figure 6. Qualitative comparison with FreeArt3D ( Chen et al., 2025a ) . We visualize the geometry of different parts using distinct colors.
Figure 7. Qualitative results on the in-the-wild images collected from the internet.
Figure 8. Qualitative results using the same images but with different masks as control signals.
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Methods \ Metric
RS- dgIoU↓
AS- dgIoU↓
RS- dcDist↓
AS- dcDist↓
RS- dCD↓
AS- dCD↓
AOR ↓
CLIP ↑
S = 1
0.1734
0.1999
0.0300
0.0966
0.0017
0.0220
0.0088
0.8903
S = 15
0.1737
0.2002
0.0320
0.0762
0.0018
0.0145
0.0090
0.8911
S = 20
0.1395
0.1395
0.0202
0.0653
0.0016
0.0139
0.0011
0.9053
Stage 1-Articulation
0.1615
0.1878
0.0302
0.0889
0.0016
0.0210
0.0068
0.8903
FM Articulation
0.3615
0.4132
0.1263
0.3019
0.0021
0.1100
0.0190
0.8803
Ours (GT Mask)
0.1395
0.1395
0.0202
0.0653
0.0016
0.0139
0.0011
0.9053
Appendix
Table 3. Ablation study with different configurations on the PartNet-Mobility dataset ( Xiang et al., 2020 ) .
Methods
Geometry
Texture
Part-level Articulation
Non-Retrieval
SINGAPO ( Liu et al., 2024a )
✓
×
✓
×
ArtFormer ( Su et al., 2025 )
✓
×
✓
✓
PhysX-3D ( Cao et al., 2025a )
✓
✓
×
✓
Articulate-Anything ( Le et al., 2025 )
×
×
✓
×
Ours
✓
✓
✓
✓
Appendix
Table 4. Comparison of representative methods and their capabilities.
Ours
ArtiAny
ArtFormer
SINGAPO
PhysX3D
PhysXAnything
FreeArt3D
~15s
~3min
~155s
~8s
~96s
~20min
~10min
Appendix
Table 5. Runtime comparison between our method and baseline methods. All measurements were conducted on a single NVIDIA A800 GPU.
Figure 9. Prompt designed for GPT-5 during the pre-processing stage to determine appropriate segmentation granularity.
Figure 10. Prompt designed for GPT-5 to perform part classification.
Figure 11. Prompt for GPT-5 in the merging stage to generate articulated part mask.
Creating interactive digital environments for gaming, robotics, and simulation relies on articulated 3D objects whose functionality emerges from their part geometry and kinematic structure. However, existing approaches remain fundamentally limited: optimization-based reconstruction methods require slow, per-object joint fitting and typically handle only simple, single-joint objects, while retrieval-based methods assemble parts from a fixed library, leading to repetitive geometry and poor generalization. To address these challenges, we introduce ArtLLM, a novel framework for generating high-quality articulated assets directly from complete 3D meshes. At its core is a 3D multimodal large language model trained on a large-scale articulation dataset curated from both existing articulation datasets and procedurally generated objects. Unlike prior work, ArtLLM autoregressively predicts a variable number of parts and joints, inferring their kinematic structure in a unified manner from the object's point cloud. This articulation-aware layout then conditions a 3D generative model to synthesize high-fidelity part geometries. Experiments on the PartNet-Mobility dataset show that ArtLLM significantly outperforms state-of-the-art methods in both part layout accuracy and joint prediction, while generalizing robustly to real-world objects. Finally, we demonstrate its utility in constructing digital twins, highlighting its potential for scalable robot learning.
Reconstructing articulated 3D objects is important for animation, gaming, and robotic simulations. Recent neural networks can estimate the articulated structure of 3D objects, but their generalization remains limited by the scarcity of annotated data for this task. To address this gap, we introduce Instruct-Particulate, a model that takes a 3D mesh together with a target kinematic specification, including part descriptions, connectivity, joint types, and optional point prompts, and predicts the corresponding kinematic part segmentation and joint motion parameters. The kinematic specification disambiguates the task and allows the model to target annotations of different granularity, thereby making it possible to use more abundant heterogeneous training data. At test time, the kinematic specification can be obtained automatically from large-scale vision-language models, so the model can be applied to any input mesh. To train our model at scale, we construct a heterogeneous dataset of more than 150,000 articulated 3D objects, extending existing publicly available collections with data obtained by partially labelling other 3D models (monolithic or already decomposed into parts) with kinematic labels by means of vision-language models. Experiments show that our model generalizes better across categories and to AI-generated meshes, enabling articulated asset reconstruction from real-world images via image-to-3D models.
Ruining Li, Yuxin Yao, Matt Zhou +5
University of Oxford · University of Cambridge · 3Nanyang Technological University
Articulated objects are fundamental for robotics, simulation of physics, and interactive virtual environments. However, recovering them from visual observations is inherently challenging, as images provide only partial and ambiguous cues about both part geometry and their underlying kinematic structure. Existing approaches typically rely on multi-stage pipelines, retrieval from asset libraries, or explicit part segmentation. We present URDF-Anything+, an end-to-end autoregressive diffusion framework that generates simulation-ready URDF models directly from a single RGB image. Conditioned on visual observations and object geometry, URDF-Anything+ operates in a structured latent space and jointly models part geometry and articulation in a unified generation process. Specifically, the model sequentially predicts each articulated part together with its associated joint parameters, while a termination token dynamically determines the number of parts. This design enables direct generation of fully executable URDFs without external retrieval or post-processing stages. Experiments on large-scale articulated object benchmarks demonstrate that URDF-Anything+ outperforms prior methods in geometric reconstruction quality, joint parameter estimation, and physical executability, while being substantially more efficient than existing multi-stage approaches. Furthermore, the generated URDFs serve as faithful digital twins, enabling the zero-shot transfer of manipulation policies trained purely in simulation.
Zhuangzhe Wu, Yue Xin, Chengkai Hou +4
Peking University · Visual Geometry Group, University of Oxford · University of Washington