Reconstructing articulated objects into high-fidelity digital twins is crucial for applications such as robotic manipulation and interactive simulation. Recent self-supervised methods using differentiable rendering frameworks like 3D Gaussian Splatting remain highly sensitive to the initial part segmentation. Their reliance on heuristic clustering or pre-trained models often causes optimization to converge to local minima, especially for complex multi-part objects. To address these limitations, we propose ArtPro, a novel self-supervised framework that introduces adaptive integration of mobility proposals. Our approach begins with an over-segmentation initialization guided by geometry features and motion priors, generating part proposals with plausible motion hypotheses. During optimization, we dynamically merge these proposals by analyzing motion consistency among spatial neighbors, while a collision-aware motion pruning mechanism prevents erroneous kinematic estimation. Extensive experiments on both synthetic and real-world objects demonstrate that ArtPro achieves robust reconstruction of complex multi-part objects, significantly outperforming existing methods in accuracy and stability.
Figures & tables
Figure 1: Existing methods like ArtGS [ 33 ] are highly sensitive to initial part segmentation, leading to inaccurate motion and geometry. Our method, ArtPro, leverages a prior-guided mobility initialization and adaptively merges mobility proposals during optimization, achieving robust reconstruction of complex multi-part articulated objects.
Figure 2: ArtPro reconstructs articulated objects from multi-view RGBD images of two states. The pipeline begins by initializing part-motion proposals through over-segmentation and a mixed-variable search. These proposals are then refined via a self-supervised optimization that updates the parts and their motion parameters using transformable 3D Gaussians. The optimization incorporates motion pruning and proposal integration operations to calibrate motions and merge proposals into coherent movable parts. Finally, a post-processing refinement stabilizes the appearance, geometry, and motion parameters of the reconstruction.
Table 34178 (5 parts)
Storage 40417 (7 parts)
Window 103238 (3 parts)
Table 23372 (5 parts)
Storage 45759 (5 parts)
Table 33116 (4 parts)
Table 34610 (6 parts)
Storage 47585 (11 parts)
All
Axis Ang
DTA [ 46 ]
46.62
20.40
4.60
33.58
6.10
2.17
34.19
31.74
22.42
ArtGS [ 33 ]
4.64
F
0.02
39.52
16.68
0.04
F
0.97
8.70 F
Ours
0.05
0.07
0.06
0.06
0.03
0.02
0.09
0.16
0.07
Axis Pos
DTA [ 46 ]
21.68
3.37
-
2.90
0.66
0.01
-
1.15
4.96
ArtGS [ 33 ]
1.42
F
-
1.46
0.05
0.00
-
0.00
0.32 F
Ours
0.00
0.00
-
0.00
0.00
0.00
-
0.00
0.00
Table 1: Quantitative evaluation on our dataset. Lower( ↓ ) is better on all metrics. We mark all joint type prediction errors and movable part count prediction errors with F . Window-103238 and Table-34610 are prismatic with no Axis Pos.
Table 25493 (4 parts)
Table 31249 (5 parts)
Storage 45503 (4 parts)
Storage 47468 (7 parts)
Oven 101908 (4 parts)
All
Axis Ang
DTA
24.35
20.62
51.18
19.07
17.83
26.61
ArtGS
1.16
0.04
0.02
0.14
0.04
0.28
Ours
0.08
0.04
0.03
0.07
0.01
0.05
Axis Pos
DTA
-
4.20
2.44
0.31
6.51
3.37
ArtGS
-
0.00
0.00
0.02
0.01
0.01
Ours
-
0.00
0.00
0.00
0.01
0.00
Table 2: ArtGS-Multi dataset [ 33 ] results. Lower( ↓ ) is better on all metrics. Table-25493 is prismatic with no Axis Pos.
Figure 3: The part initialization and reconstruction results of ArtGS [ 33 ] and ours on two-part objects. Although our approach doesn’t merge the two disadjacent movable parts (second row), we still obtain accurate motion and geometry reconstruction.
Figure 4: The reconstructed articulated objects in different motion states ( t={0,0.5,1} ) and their part-motion structures.
Figure 5: The intermediate results during our adaptive proposal integration optimization. We show the optimized Gaussians (with their center points) and the estimated motions before the proposal integration operation at each cycle.
Figure 6: Ablation study results. To compare with our OverSeg initialization, we use DBSCAN [ 7 ] as our clustering implementation, denoted Vanilla . We further add motion initialization to Vanilla and OverSeg baseline, denoted w/MoI . Finally, we removed the pruning strategy from Full ( w/oPrune ) to verify its significant.
Figure 7: Reconstructed results of real-world articulated objects.
Case
OverSeg
MoI
Merge
Prune
CD-s
CD -m
CD-w
#1
1.46
9.27
0.93
#2
✓
1.04
6.97
0.67
#3
✓
✓
0.87
5.01
0.71
#4
✓
✓
✓
0.85
4.57
0.67
Full
✓
✓
✓
✓
0.84
3.63
0.65
Table 3: Quantitative evaluation of the ablation study. Lower( ↓ ) is better on all metrics. To evaluate the results with arbitrary numbers of movable parts, we compute CD between the merged movable results and GT movable parts, denoted as over CD -m.
Appendix figures & tables11 assets
Supplementary material from the paper’s appendix.
Appendix
Synthetic Objects
Real Objects
FoldChair
Fridge
Laptop
Oven
Scissor
Stapler
USB
Washer
Blade
Storage
All
Fridge
Storage
All
Axis Ang
PARIS* [ 25 ]
15.79
2.93
0.03
7.43
16.62
8.17
0.71
0.71
41.28
0.03
9.37
1.90
30.10
16.00
ArticulatedGS [ 11 ]
6.20
0.14
15.55
0.00
0.07
0.08
0.16
0.03
0.23
0.04
2.25
4.06
48.56
26.31
ArticulatedGS*
0.02
0.12
0.03
0.13
0.18
0.13
0.12
0.17
0.04
0.04
0.10
64.81
29.44
47.12
DTA [ 46 ]
0.03
0.09
0.07
0.22
0.10
0.07
0.11
0.36
0.20
0.09
0.13
2.08
13.64
7.86
ArtGS [ 33 ]
0.01
0.03
0.01
0.01
0.05
0.01
0.04
0.02
0.03
0.01
0.02
2.09
3.47
2.78
Appendix
Table 4: Results on the PARIS [ 25 ] dataset, including both synthetic and real data. Methods marked with an asterisk () denote versions with added depth supervision for fair comparison. Specifically, ArticulatedGS and PARIS* are trained with an additional depth loss.
Figure 8: Additional qualitative results on PARIS dataset, including the initializations and the final reconstructions. We show the Gaussians with their center points for a better visualization of their segmentation and motion parameters.
Figure 9: Additional reconstruction results on ArtGS-Multi dataset . We render the reconstructed articulated objects in different motion states ( t∈{0,0.5,1} and their motion structures for each result.
Figure 10: Additional qualitative results on ArtGS-Multi dataset , including the initializations and the final reconstructions. We show the Gaussians with their center points for a better visualization of their segmentation and motion parameters.
Figure 11: Additional reconstruction results on our dataset . We render the reconstructed articulated objects in different motion states ( t∈{0,0.5,1} and their motion structures for each result.
Figure 12: Additional qualitative results on our dataset , including the initializations and the final reconstructions. We show the Gaussians with their center points for a better visualization of their segmentation and motion parameters.
Method
PSNR
SSIM
LPIPS
ArtGS
36.10
0.980
0.037
Ours
48.02
0.998
0.003
Appendix
Table 5: Visual quality on our dataset.
Case
CD-s
CD -m
CD-w
w/o PM
0.85
4.26
0.68
w/o MoI
0.89
5.62
0.70
w/o MoI&PI
0.87
5.81
0.71
Full
0.84
3.63
0.65
Appendix
Table 6: Additional ablation study on our dataset.
Method
Time (min)
GPU Memory (GB)
DTA [ 46 ]
53.28 ±23.78
10.58 ±2.45
ArtGS [ 33 ]
10.41 ±0.89
2.94 ±0.04
Ours
16.93 ±2.37
2.69 ±0.23 *
Appendix
Table 7: Computation overhead comparison on our multi-part dataset. Time and memory are reported as mean ± std.
Figure 13: Failure cases. We illustrate failure cases of low-quality part boundaries. For low-quality RGBD images input, our method difficult to reconstruct the part boundaries.
Reconstructing articulated objects with multiple movable parts is essential for understanding object structure and enabling physical interaction. However, this reconstruction task poses significant challenges due to the entanglement of geometry, appearance, and motion parameters during optimization. Existing methods rely primarily on photometric supervision, which commonly fails to disentangle these interdependent components, resulting in poor part decomposition with blurred boundaries and geometric artifacts. To address this limitation, we introduce StructureGS, a reconstruction framework for articulated objects that integrates structure-aware guidance into 3D Gaussian Splatting. Our approach leverages oriented bounding boxes of object parts to enforce two key structural properties: spatial coherence, which constrains each part's geometry to remain compact and spatially coherent within its designated region, and structural connectivity, which enforces physically plausible contact relationships between adjacent parts. These properties are realized through structure-aware losses that inject explicit structural constraints into the optimization process. Extensive experiments demonstrate that our method achieves state-of-the-art performance in articulated object reconstruction, producing high-quality results with well-defined part geometries.
Reconstructing articulated objects is essential for building digital twins of interactive environments. However, prior methods typically decouple geometry and motion by first reconstructing object shape in distinct states and then estimating articulation through post-hoc alignment. This separation complicates the reconstruction pipeline and restricts scalability, especially for objects with complex, multi-part articulation. We introduce a unified representation that jointly models geometry and motion using articulated 3D Gaussians. This formulation improves robustness in motion decomposition and supports articulated objects with up to 20 parts, significantly outperforming prior approaches that often struggle beyond 2--3 parts due to brittle initialization. To systematically assess scalability and generalization, we propose MPArt-90, a new benchmark consisting of 90 articulated objects across 20 categories, each with diverse part counts and motion configurations. Extensive experiments show that our method consistently achieves superior accuracy in part-level geometry reconstruction and motion estimation across a broad range of object types. We further demonstrate applicability to downstream tasks such as robotic simulation and human-scene interaction modeling, highlighting the potential of unified articulated representations in scalable physical modeling.
Reconstructing articulated 3D objects from a single image requires jointly inferring object geometry, part structure, and motion parameters from limited visual evidence. A key difficulty lies in the entanglement between motion cues and object structure, which makes direct articulation regression unstable. Existing methods address this challenge through multi-view supervision, retrieval-based assembly, or auxiliary video generation, often sacrificing scalability or efficiency. We present MonoArt, a unified framework grounded in progressive structural reasoning. Rather than predicting articulation directly from image features, MonoArt progressively transforms visual observations into canonical geometry, structured part representations, and motion-aware embeddings within a single architecture. This structured reasoning process enables stable and interpretable articulation inference without external motion templates or multi-stage pipelines. Extensive experiments on PartNet-Mobility demonstrate that OM achieves state-of-the-art performance in both reconstruction accuracy and inference speed. The framework further generalizes to robotic manipulation and articulated scene reconstruction.