EmbodiedGen – A Generative 3D World Engine for Embodied Intelligence Applications

What is EmbodiedGen?

EmbodiedGen is a generative 3D world engine and toolkit designed for embodied intelligence (Embodied AI) applications. It enables the rapid creation of high-quality, low-cost, and physically plausible 3D assets and interactive environments, assisting researchers and developers in building test environments for embodied agents.

EmbodiedGen features multiple modules, including image-to-3D, text-to-3D, texture generation, articulated object generation, scene and layout construction, and more. It supports the generation of everything from simple objects to complex, realistic environments. The generated 3D assets are compatible with robotic simulation and support formats like URDF (Unified Robot Description Format), making it a powerful tool for embodied AI research.

Key Features of EmbodiedGen

• Image-to-3D Conversion: Converts input images into 3D assets with physically plausible structures.

• Text-to-3D Generation: Creates a variety of 3D objects based on textual descriptions, with diverse geometric shapes and styles.

• Texture Generation: Adds visually rich textures to 3D meshes.

• Complex Scene Construction: Enables the generation of realistic and structured 3D environments—from single objects to full scenes—compliant with real-world scales and URDF formats.

• Intelligent Layout Generation: Automatically arranges objects within a scene, facilitating downstream training and evaluation tasks.

• Physical Property Support: Generated 3D assets feature watertight geometry and physically realistic attributes, ready for direct use in robotic simulations and formal description standards.

Technical Principles Behind EmbodiedGen

• Use of Generative AI: Powered by generative AI, EmbodiedGen can synthesize 3D models from both images and text descriptions.

• Multi-Module Collaboration: The system integrates six core modules—image-to-3D, text-to-3D, texture generation, articulated object generation, scene generation, and layout generation. These modules work together to create diverse 3D environments ranging from simple objects to rich scenes.

• Physical Realism and Real-World Scaling: All generated assets have watertight geometry and physically sound properties, ensuring compatibility with robotics simulation standards like URDF.

• Dynamic Environment Generation: EmbodiedGen supports dynamic environments that can evolve in real time based on the AI agent’s actions.

Project Links

• Official Website: https://horizonrobotics.github.io/robot_lab/embodied_gen/index.html

• GitHub Repository: https://github.com/HorizonRobotics/EmbodiedGen

• arXiv Paper: https://arxiv.org/pdf/2506.10600

Application Scenarios of EmbodiedGen

• Robotics Simulation & Training: EmbodiedGen produces physically accurate, real-world scale 3D assets ready for integration with URDF and other robotics simulation formats.

• Autonomous Driving & Drones: Dynamic 3D environments generated by EmbodiedGen simulate complex roads and terrains, providing robust training grounds for autonomous driving systems and drones to better adapt to real-world conditions.

• Virtual Social Interaction: Through VR devices, users can control avatars to engage in social activities, meetings, and collaboration within virtual environments.

• Healthcare & Rehabilitation: EmbodiedGen’s 3D environments can be used in medical and rehabilitation simulations and training scenarios—such as practicing surgical operations in a safe, virtual space.