Towards Physically Grounded Articulated 3D Assets: A Multi-Agent Pipeline and Benchmark

Xudong Cai^1#*Ruijie Zhu^1*Haoliang Shang^1,2*Ruining Li^3*

Marc Pollefeys^2,4Shangzhe Wu⁵Andrea Vedaldi³Chuanxia Zheng¹

¹ PVG, Nanyang Technological University² ETH Zurich³ VGG, University of Oxford⁴ Microsoft⁵ University of Cambridge

# Project lead · * Core technical contribution

Paper Code Explore assets Interactive demo Benchmark

Office · Original photo taken in Chuanxia Zheng’s office

Explore this room

Explore the rooms

Artisan-generated articulated objects assembled into an office, a laboratory, and a kitchen, with examples of per-part materials and masses.

Interactive demo

Physical validation

94.6%

Lower single-joint collision rate

vs. Articulate-Anything · 31.2% → 1.7%

14.0%

Higher rest-state stability

vs. Codex · 86% → 98%

105.2%

Higher first-choice preference

vs. Codex · 21.1% → 43.3%

Relative improvements over the best non-Artisan baseline for each metric. Physics scores use Artisan (GPT-6 Astra).

Collision

Red boxes highlight incorrect motion detected by physical collision checks.

Stability

Robot manipulation

Red border: execution failed.

Full video · 5:33 ↗

Contributions

Multi-agent generation

Design, Coding, and Review agents separate planning, construction, and independent assessment.

Physical validation

Simulation detects collisions and instability, then guides repairs to geometry and physical parameters.

A unified benchmark

More than 10 models evaluated on appearance, articulation, physics, robot tasks, and human preference.

Read the abstract

In this paper, we study physically grounded articulated 3D asset generation—moving the goal from reproducing how objects look to capturing how they move and function. Inspired by the literature on procedural and agentic generation, we present Artisan, a system that extends the latest method Articraft by introducing a multi-agent pipeline with physical simulation feedback, enabling comprehensive attributes beyond geometry, appearance, and physically plausible articulation, such as collision-free motion, stability under gravity, and accurate mass and material properties. Furthermore, we construct a unified benchmark spanning four key aspects: spatial fidelity, physical plausibility, downstream task performance, and overall quality. The benchmark encompasses 100 reference assets from ArtVIP and LightWheel, 50 stability prompts, and 50 ABO objects with reference masses, with protocols evaluating various metrics. Through extensive evaluation of more than 10 representative models, Artisan achieves competitive spatial fidelity while attaining the lowest collision rates and highest stability rates among the evaluated methods. Together, our system and benchmark establish insights for advancing 3D generation towards physically plausible, functionally usable assets.

Method

Text or image input flows through Design, Coding, and Review agents. Collision and stability reports guide revisions before exporting an articulated asset with physics and action annotations. — Design, generate, validate, and review. Figure 2 · Click to enlarge ↗

Benchmark and results

4 aspects, 5 datasets, and more than 20 metrics

Spatial fidelity

Physical plausibility

Downstream tasks

Overall quality

Detailed metrics & rankingsShow details Hide details

Detailed results

Table 2 ↗

Rank by

Lower is better ↓

ArtVIP · collision under motion

Single-joint collision · 13 evaluated configurations
Rank	Method / backbone	Single-joint collision ↓	Double-joint collision ↓	Multi-joint collision ↓
1	ArtisanGPT-6 Astra	1.7%	3.7%	13.0%
2	ArtisanGPT-5.6 Sol	1.9%	4.8%	16.4%
3	ArtisanGPT-5.5	3.5%	6.4%	12.3%
4	Articulate-Anything	31.2%	39.2%	32.8%
5	CodexGPT-6 Astra	32.4%	47.0%	63.7%
6	ArtisanGPT-5.5 · without physics	37.5%	47.3%	77.6%
7	ClaudeFable 5.1	39.9%	51.7%	64.7%
8	ArticraftGPT-5.5	44.5%	57.5%	79.4%
9	ArticraftGPT-6 Astra	45.5%	60.7%	78.8%
10	AgyGemini-Flash-3.8	46.3%	65.6%	81.5%
11	PhysX-Anything	58.4%	77.6%	81.3%
12	Instruct-Particulate	70.4%	90.4%	98.7%
13	URDF-Anything+	84.6%	99.3%	100.0%
—	NeRF2Physics	—	—	—
—	PUGS	—	—	—

Best scoreSecond-best score= Shared rank— Not reported

All 15 rows from Table 2, including backbone variants and the physics ablation. An em dash means the paper reports no result; that row is unranked for the selected metric.

Ranks use the paper’s reported values for the selected metric; ties share a rank. Each dimension is evaluated separately. Click a column heading to rank by that metric.

Overview charts: spatial and physical results use GPT-6 Astra for Artisan, Articraft, and Codex; spatial results are on ArtVIP. Downstream results compare training compositions with a π0.5 policy on held-out native assets. Human preference follows visual inspection and physical interaction.

Evaluation protocols and dataset composition: Sections 4.1–4.2 and Appendix B ↗

Takeaways

Agentic generation methods can rival or even surpass state-of-the-art task-specific methods.

Physical validation is the key to generating a ready-to-use asset.

Multi-agent design and independent review help eliminate the self-evaluation bias.

Adding generated assets to native robot training data improves generalization to unseen objects.

Citation

@misc{cai2026artisan,
  title  = {Towards Physically Grounded Articulated 3D Assets: A Multi-Agent Pipeline and Benchmark},
  author = {Cai, Xudong and Zhu, Ruijie and Shang, Haoliang and Li, Ruining and Pollefeys, Marc and Wu, Shangzhe and Vedaldi, Andrea and Zheng, Chuanxia},
  year   = {2026},
  eprint = {2610.13947},
  archivePrefix = {arXiv},
  url    = {https://arxiv.org/abs/2610.13947}
}