Articraft: An Agentic System for Scalable Articulated 3D Asset Generation

Forum topic · 小凯 · 2026-05-16

Summary

Articraft is an agentic system that uses large language models (LLMs) to generate articulated 3D assets at scale, addressing the bottleneck of scarce training data for articulated object understanding. The approach reduces 3D asset generation to program writing: an LLM writes code against a domain-specific SDK to define parts, compose geometry, specify joints, and author validation tests. A harness exposes a restricted workspace and interface, validates generated assets, and returns structured feedback, freeing the LLM from low-level details like writing URDF files or managing complex software environments. The authors show that Articraft produces assets of higher quality than state-of-the-art articulated asset generators and general coding agents. Using the system, they build Articraft-10K, a curated dataset of over 10K articulated assets spanning 245 categories, and demonstrate its utility for training articulated asset models and downstream applications such as robot simulation and virtual reality. Paper: arXiv 2505.08631.

Paper Overview

Field: Computer Vision (CV)
Authors: Matt Zhou, Ruining Li, Xiaoyang Lyu
arXiv: 2505.08631

Abstract (Original)

A bottleneck in learning to understand articulated 3D objects is the lack of large and diverse datasets. In this paper, we propose to leverage large language models (LLMs) to close this gap and generate articulated assets at scale. We reduce the problem of generating an articulated 3D asset to that of writing a program that builds it. We then introduce a new agentic system, Articraft, that writes such programs automatically. We design a programmatic interface and harness to help the LLM do so effectively. The LLM writes code against a domain-specific SDK for defining parts, composing geometry, specifying joints, and writing tests to validate the resulting assets. The harness exposes a restricted workspace and interface to the LLM, validates the resulting assets, and returns structured feedback. In this way, the LLM is not distracted by details such as writing URDF files or managing complex software environments. We show that this approach yields assets of higher quality than state-of-the-art articulated asset generators and general coding agents. With Articraft, we construct Articraft-10K, a curated dataset of over 10K articulated assets spanning 245 categories, and demonstrate its utility in training articulated asset models and downstream applications such as robot simulation and virtual reality.

Key Points

Problem: Lack of large, diverse datasets limits learning to understand articulated 3D objects.
Idea: Frame articulated asset generation as program writing, automated via an LLM-based agentic system.
Design: A domain-specific SDK plus a harness with a restricted workspace; the LLM defines parts, composes geometry, specifies joints, and writes validation tests, receiving structured feedback from asset validation.
Results: Higher-quality assets than state-of-the-art articulated generators and general coding agents.
Dataset: Articraft-10K — over 10K articulated assets across 245 categories, useful for training articulated asset models, robot simulation, and VR.

--- *Auto-collected 2026-05-16.*