ArtiPG++: Towards Efficient Procedural Generation of Articulated Object Data

Jiude Wei, Nange Wang, Longfei Xu, Yuxuan Li, Cewu Lu, Jianhua Sun

Keywords: Articulated Object, Procedural Generation, Articulated Object Manipulation, Robotics

Abstract: To leverage deep learning in advancing vision perception and embodied intelligence, an extensive number of high-quality and richly annotated 3D articulated objects is essential. However, current methods for collecting articulated objects and their annotations are either based on human effort or physics simulators, which are difficult to scale up, posing challenges to the collection of large-scale and richly annotated articulated objects. In such context, procedural generation has recently gained attention in articulated object synthesis. However, it still faces challenges such as reliance on external assets and the complexity of designing procedural rules. To this end, we propose ArtiPG++, a highly efficient framework for synthesizing articulated objects with rich annotations, featuring three key advantages: 1) asset-free spatial structure synthesis via procedural rules, 2) labor-free synthesis of realistic geometric details, along with precise and diverse annotations, and 3) easy expansion to new object categories, with a ready-to-use tool for convenient synthesis. ArtiPG++ currently supports the procedural synthesis for 39 common object categories, and requires only a few hours to develop procedural generation rules for novel categories, which is a one-time effort for infinite objects synthesis. We conduct extensive evaluations on the objects and annotations synthesized by ArtiPG++, through both direct comparisons in terms of diversity and distribution, as well as performance in downstream tasks. Please refer to the appendix for more details, analysis, discussions and code implementation.

Primary Area: applications to robotics, autonomy, planning

Submission Number: 1342