EmbodiedGen - A generative 3D world engine for embodied intelligence applications

What is EmbodiedGen?

EmbodiedGen is a generative 3D world engine and toolkit for Embodied AI applications. It can quickly generate high-quality, low-cost, and physically plausible 3D assets and interactive environments, helping researchers and developers build test environments for embodied agents. EmbodiedGen includes multiple modules, such as generating 3D models from images or text, texture generation, jointed object generation, scene and layout generation, supporting the creation of everything from simple objects to complex scenes. The generated 3D assets can be directly used for robot simulation and in URDF format, providing powerful tool support for embodied AI research.

The main functions of EmbodiedGen

• Image to 3D conversionIt can generate physically plausible 3D assets from input images.
• Text to 3D generationGenerates 3D assets with various geometric shapes and styles based on text descriptions.
• Texture generation functionGenerate visually rich textures for 3D meshes.
• Complex scene constructionIt supports the creation of everything from simple objects to complex scenes, and can generate high-quality 3D assets with real-world scale and conforming to the Uniform Robot Description Format (URDF).
• Intelligent layout generationIt provides intelligent layout generation capabilities to support downstream tasks in training and evaluation.
• Physical property supportThe generated 3D assets have sealed geometry and physically reasonable properties, and can be directly applied to robot simulation and description formats.

EmbodiedGen's technical principles

• Applications of Generative AIEmbodiedGen is based on generative AI technology and can generate 3D models from images or text descriptions.
• Multi-module collaborative workEmbodiedGen comprises six key modules: image-to-3D, text-to-3D, texture generation, jointed object generation, scene generation, and layout generation. These modules work together to generate diverse 3D worlds, ranging from simple objects to complex scenes.
• Physical Realism and Real-World ScaleThe generated 3D assets have sealed geometry and physically reasonable properties, and can be directly applied to robot simulation and description formats such as URDF (Unified Robot Description Format).
• Dynamic environment generationEmbodiedGen's generation environment is dynamic, capable of generating and modifying the environment in real time based on the AI's behavior.

EmbodiedGen's project address

• Project official website: https://horizonrobotics.github.io/robot_lab/embodied_gen/index.html
• Github repositoryhttps://github.com/HorizonRobotics/EmbodiedGen
• arXiv technical paper: https://arxiv.org/pdf/2506.10600

Application scenarios of EmbodiedGen

• Robot Simulation and TrainingEmbodiedGen can generate 3D assets with physical plausibility and real-world scale, which can be directly applied to robot simulation and description formats such as URDF (Unified Robot Description Format).
• Autonomous driving and dronesEmbodiedGen generates dynamic 3D environments that can be used for simulation training of autonomous driving and drones. By simulating complex road and terrain conditions, it helps autonomous driving systems and drones better adapt to real-world scenarios.
• Virtual socialUsers can control virtual avatars through VR devices to conduct social activities, meetings, and other activities.
• Medical treatment and rehabilitationEmbodiedGen generates 3D environments that can be used for simulation and training in the medical and rehabilitation fields. This includes simulation training for surgical procedures within a virtual environment.