EmbodiedGen is a generative 3D world engine designed for embodied intelligence research, enabling the creation of diverse, interactive 3D environments with high-quality assets and plausible physics. It addresses the generalization challenges in embodied AI by offering modules for image-to-3D, text-to-3D, texture generation, articulated object generation, scene generation, and layout generation.

How It Works

EmbodiedGen leverages a suite of generative AI models to construct 3D worlds. Key components include image-to-3D for generating URDFs from single images, text-to-3D using models like SD3.5 and Kolors for asset creation, and texture generation for applying rich visual details. The system also supports 3D scene generation and aims to incorporate articulated object and interactive layout generation, providing a comprehensive toolkit for building complex simulated environments.

Quick Start & Requirements

• Installation: Clone the repository, checkout v0.1.2, initialize submodules, and run bash install.sh basic. A Docker image is also available.
• Prerequisites: Python 3.10.13, CUDA. GPT agent setup requires an API key for GPT-4o or Qwen2.5-VL.
• Resources: Model checkpoints are downloaded on first run. Scene generation can take ~30 minutes per scene.
• Links: Docker Documentation

Highlighted Details

• Image-to-3D generates physically plausible 3D asset URDFs from single images.
• Text-to-3D supports multiple models (SD3.5, Kolors) and languages for asset generation.
• 3D Scene Generation can create scenes from text prompts using scene3d-cli.
• Includes a "Real-to-Sim Digital Twin" module for developers.

Maintenance & Community

The project cites an arXiv paper from 2025, indicating recent development. Acknowledgements list numerous foundational projects and models, suggesting a strong reliance on existing research.

Licensing & Compatibility

Licensed under the Apache License 2.0, permitting commercial use and integration with closed-source projects.

Limitations & Caveats

Articulated Object Generation and Layout Generation modules are marked as "Coming Soon." The project relies on external GPT models, requiring API key configuration.