URDF-Anything+: End-to-End Generation for Simulation-Ready Articulated Assets

Zhuangzhe Wu; Yue Xin; Chengkai Hou; Minghao Chen; Yaoxu Lyu; Jieyu Zhang; Shanghang Zhang

cs.ROMar 14, 2026

URDF-Anything+: End-to-End Generation for Simulation-Ready Articulated Assets

Authors: Zhuangzhe Wu, Yue Xin, Chengkai Hou, Minghao Chen, Yaoxu Lyu, Jieyu Zhang, Shanghang Zhang

Organizations: Peking University · Visual Geometry Group, University of Oxford · University of Washington

Abstract

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.

Figures & tables

Figure 1: Comparison between previous methods and our method. In (a), we summarize previous generation paradigms for articulated objects, including different segmentation approaches and joint parameter generation methods. In (b), we briefly illustrate our method and its advantages.

Panel of Figure 2 — Figure 2: Overview of URDF-Anything+. (a) The autoregressive generation pipeline that produces 3D parts together with their joint specifications and assembles them into a complete URDF. (b) Structure of the DiT backbone and the geometry / articulation parameter decoders.

Figure 3: Qualitative Results on the Test Set of Our Dataset. Compared with other methods, URDF-Anything+ generates high-quality 3D assets with more accurate geometry and articulation.

Method	Parts			Whole Object			Joint Accuracy			Time / Part (s) $\downarrow$
Method	IoU $\uparrow$	F-Score $\uparrow$	CD $\downarrow$	IoU $\uparrow$	F-Score $\uparrow$	CD $\downarrow$	Axis Err. $\downarrow$	Origin Err. $\downarrow$	Limit Err. $\downarrow$	Time / Part (s) $\downarrow$
URDFormer Chen et al. (2024)	0.202	0.389	0.354	0.243	0.413	0.286	1.071 ±0.497	0.504 ±0.331	0.882 ±0.679	41.88
Articulate-Anything Le et al. (2025)	0.563	0.452	0.064	0.627	0.510	0.053	0.230 ±0.505	0.267 ±0.357	1.108 ±0.676	380.57
SINGAPO Liu et al. (2025a)	0.382	0.434	0.165	0.462	0.481	0.132	0.670 ±0.472	0.404 ±0.297	0.817 ±0.525	52.30
PhysX-Anything Cao et al. (2025b)	0.784	0.697	0.076	0.813	0.715	0.068	0.191 ±0.337	0.103 ±0.192	0.532 ±0.423	578.38
URDF-Anything+ (Ours)	0.879	0.721	0.033	0.930	0.742	0.009	0.129 ±0.404	0.062 ±0.096	0.225 ±0.440	34.70

Table 1: Main Results. Geometry reconstruction quality (Parts and Whole Object), joint parameter prediction accuracy, and average inference time per generated part. We report IoU and F-Score (higher is better,

\uparrow

), CD (lower is better,

\downarrow

), joint Axis (rad), Origin (m), Limit (rad) errors (lower is better,

\downarrow

), and per-part inference time (s, lower is better,

\downarrow

). All reported times include geometry decoding and URDF assembly, and are measured on the same hardware with identical input configurations across methods.

Figure 4: Qualitative Results on In-the-wild Images . Compared with other methods, URDF-Anything+ generates high-quality 3D assets with more accurate geometry and articulation.

Objectives	Variation	Geometry Quality						Joint Accuracy
		Parts			Whole Object			Axis Err. $\downarrow$	Origin Err. $\downarrow$	Limit $\downarrow$
		IoU $\uparrow$	F-Score $\uparrow$	CD $\downarrow$	IoU $\uparrow$	F-Score $\uparrow$	CD $\downarrow$	Axis Err. $\downarrow$	Origin Err. $\downarrow$	Limit $\downarrow$
Input Modality	Image Only	0.627	0.560	0.106	0.682	0.634	0.098	0.349	0.197	0.540
Input Modality	Image + 3D Guid.(Ours)	0.823	0.740	0.028	0.841	0.766	0.009	0.182	0.091	0.281
Link Ordering	Random Order	0.630	0.686	0.145	0.710	0.704	0.072	0.203	0.168	0.442
Link Ordering	Order by Joint XYZ (Ours)	0.805	0.742	0.096	0.816	0.753	0.034	0.120	0.102	0.323

Table 2: Ablation Studies. We evaluate the impact of input modalities (top) and link ordering strategies (bottom) on geometry reconstruction and joint prediction accuracy.

Panel of Figure 5 — Figure 5: Real-Follow-Sim deployment. (a) Pipeline for constructing the simulation digital twin from real RGB-D observations. (b) The real-world tasks used for evaluation.

Comparison of deployment paradigms			Comparison of URDF generation methods (under Real-Follow-Sim)			Comparison of training data sources
Method	Close Laptop	Close Drawer	Method	Close Laptop	Close Drawer	Method	Close Laptop	Close Drawer
Sim-to-Real	80	75	Articulate-Anything	75	55	Real-Only	85	75
Real-Follow-Sim (Ours)	100	90	URDF-Anything+ (Ours)	100	90	Simulation-Only (Ours)	100	90

Table 3: Real-World performance evaluation. We compare our method across different deployment paradigms, URDF generation baselines, and training data sources. Success rates (%) are reported.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 6 — Figure 6: Failure case caused by single-layer meshes. From left to right: the input mesh, sampled surface points with unreliable normals, and the failed reconstruction after encoding and decoding.

Panel of Figure 7 — Figure 7: Effect of mesh thickening. From left to right: the thickened mesh, sampled surface points with well-defined normals, and the decoded reconstruction after encoding and decoding. Mesh thickening stabilizes normal estimation and enables reliable latent encoding.

Panel of Figure 8 — Figure 8: The visualization of link order setting

Panel of Figure 9 — Figure 9: Examples of rendered images generated by Qwen-Image-Edit

Figure 10: Generation Results of Microwaves.

Figure 11: Generation Results of Refrigerators.

Figure 12: Generation Results of Faucets.

Figure 13: Generation Results of Dishwashers.

Articulated 3D Asset Generation

See all 16 papers

PAct: Part-Decomposed Single-View Articulated Object Generation
Feb 16, 2026Qingming Liu, Xinyue Yao, Shuyuan Zhang +4Articulated 3D Asset Generation 3D Asset Generation
The Chinese University of Hong Kong, Shenzhen, China · DexForce Technology, China
ArtLLM: Generating Articulated Assets via 3D LLM
Mar 1, 2026Penghao Wang, Siyuan Xie, Hongyu Yan +4Articulated 3D Asset Generation 3D Asset Generation
ShanghaiTech University · Tencent Hunyuan · HKUST
SIMART: Decomposing Monolithic Meshes into Sim-ready Articulated Assets via MLLM
Mar 24, 2026Chuanrui Zhang, Minghan Qin, Yuang Wang +3Articulated 3D Asset Generation 3D Asset Generation
ByteDance Seed · NTU

Feb 16, 2026J/K move · Enter open · S save

PAct: Part-Decomposed Single-View Articulated Object Generation

Qingming Liu, Xinyue Yao, Shuyuan Zhang, Yueci Deng+3

The Chinese University of Hong Kong, Shenzhen, China · DexForce Technology, China

Articulated 3D Asset Generation 3D Asset Generation Part-Level 3D Generation Single-View 3D Reconstruction

Open paper AskPDF arXiv

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.

Abstract

Articulated 3D Asset Generation

Explore similar work

PAct: Part-Decomposed Single-View Articulated Object Generation

ArtLLM: Generating Articulated Assets via 3D LLM

SIMART: Decomposing Monolithic Meshes into Sim-ready Articulated Assets via MLLM