Articraft

Articraft is an open-source agentic system that transforms the creation of articulated 3D assets into a programmatic, code-generation workflow powered by large language models. Built by Matt Zhou and collaborators from Oxford's Visual Geometry Group, it is designed for large-scale dataset generation and bypasses heavyweight manual 3D modeling tools. The project is accompanied by a research paper (arXiv:2605.15187) and is licensed under Apache 2.0.

What It Is

Articraft sits at the intersection of generative AI and 3D asset creation. Rather than requiring artists to manually rig and model objects, it uses LLMs to generate Python-based model definitions (model.py files) that describe articulated objects with semantic parts, robust geometry, and physical joints. The system is engineered specifically for producing large, diverse datasets of articulated 3D models — the kind needed to train and evaluate robotics, simulation, and computer vision systems.

How the Agentic Workflow Operates

The core workflow centers on a CLI command (articraft generate) that accepts a natural-language prompt and produces a structured dataset record. The system defaults to high-capability models (e.g., GPT or Gemini variants) and supports configurable cost caps and model selection. Key workflow steps include:

• Generate: Produce a new articulated asset from a text prompt
• Fork: Modify an existing record by branching it into a child record, leaving the parent unchanged
• Hydrate: Pull Git LFS-stored record payloads on demand by record ID, category, date range, or batch
• View: Browse and inspect the dataset through a local React frontend and API server

External AI agents (Claude Code, Codex, Cursor) can also contribute to the dataset by following a documented workflow, enabling crowdsourced dataset growth without requiring direct API key access.

Architecture and Dataset Design

Articraft stores dataset records in a code-first format under data/records/, tracked via Git LFS with lazy hydration. A data/records_index.jsonl file indexes all records so the viewer can browse metadata without downloading full payloads. Each record contains a model.py file that is executed as Python code to compile and inspect the asset — the project notes this carries a security implication and advises running only trusted sources.

The project structure is documented across several guides covering architecture, record editing, batch dataset generation, and contribution standards.

Contribution and Data Licensing

A stated goal of Articraft is crowdsourcing a large, diverse dataset of articulated 3D models. Contributions via CLI, batch processing, or external AI agents are welcomed through pull requests. The project specifies that all contributed data is released under Creative Commons Attribution 4.0 International (CC-BY 4.0), and contributors agree their submissions may be used to build and evaluate machine learning models distributed publicly.

Current Status

The repository was created in March 2026 and last updated in May 2026, with 891 stars and 105 forks reported on GitHub. The project supports Python 3.11 and 3.12 (3.13+ is not currently supported). CI is active via GitHub Actions. The accompanying research paper is available as an arXiv preprint (arXiv:2605.15187).