Abstract
Constructing a physically realistic and accurately scaled simulated 3D worldis crucial for the training and evaluation of embodied intelligence tasks. Thediversity, realism, low cost accessibility and affordability of 3D data assetsare critical for achieving generalization and scalability in embodied AI.However, most current embodied intelligence tasks still rely heavily ontraditional 3D computer graphics assets manually created and annotated, whichsuffer from high production costs and limited realism. These limitationssignificantly hinder the scalability of data driven approaches. We presentEmbodiedGen, a foundational platform for interactive 3D world generation. Itenables the scalable generation of high-quality, controllable andphotorealistic 3D assets with accurate physical properties and real-world scalein the Unified Robotics Description Format (URDF) at low cost. These assets canbe directly imported into various physics simulation engines for fine-grainedphysical control, supporting downstream tasks in training and evaluation.EmbodiedGen is an easy-to-use, full-featured toolkit composed of six keymodules: Image-to-3D, Text-to-3D, Texture Generation, Articulated ObjectGeneration, Scene Generation and Layout Generation. EmbodiedGen generatesdiverse and interactive 3D worlds composed of generative 3D assets, leveraginggenerative AI to address the challenges of generalization and evaluation to theneeds of embodied intelligence related research. Code is available athttps://horizonrobotics.github.io/robot_lab/embodied_gen/index.html.