ArtiWorld: LLM-Driven Articulation of 3D Objects in Scenes
Abstract
Building interactive simulators and scalable robot-learning environments requires a large number of articulated assets. However, most existing 3D assets used in simulation are rigid, and manually converting them into articulated objects with executable kinematic structures is highly labor-intensive, time-consuming, and expensive. This limitation significantly restricts the scalability of simulator construction and downstream embodied AI research. To address this challenge, we present ArtiWorld, a scene-aware pipeline that automatically localizes rigid objects in 3D environments and converts them into interactive URDF models. At the core of this pipeline is Arti4URDF, a framework that integrates 3D point cloud geometry, the prior knowledge of a large language model (LLM), and a URDF-oriented prompt design to infer articulated structures and generate executable articulated assets while preserving the original 3D shape. By explicitly modeling object parts, geometric relations, and articulation priors, our method enables direct conversion from rigid 3D assets to robot-ready articulated objects. We evaluate our method on simulated articulated objects, full-scene assets, and real-world scanned data. Experimental results show that ArtiWorld achieves state-of-the-art performance and consistently outperforms existing methods in articulation prediction quality and asset usability. Furthermore, the generated assets can be directly deployed in simulation environments and even can be generalized to real-world. Code and data will be released.