Building 3D objects that move: an AI system for articulated assets

Creating diverse datasets of articulated 3D objects (objects with moving parts) is currently labor-intensive, limiting research in embodied AI and robotics simulation. Articraft solves this by treating asset generation as code synthesis: an LLM writes programs against a domain-specific SDK to define parts, geometry, joints, and tests. A structured harness validates outputs and provides feedback without exposing the LLM to low-level file formats or environment complexity. The system produces higher-quality assets than existing articulated-asset generators and general coding agents. The team released Articraft-10K, a curated dataset of over 10,000 articulated objects across 245 categories, demonstrating utility for training articulated-object models and downstream robotics simulation and VR tasks.