HorizonRobotics/EmbodiedGen

What it solves

EmbodiedGen is a generative 3D world engine designed to bridge the gap between high-level intent and executable simulation environments. It allows researchers and developers to quickly create physically plausible 3D assets, large-scale indoor scenes, and interactive, task-driven worlds that are immediately deployable across multiple major robotics simulators without manual adaptation.

How it works

The engine compiles language, images, and edit commands into simulation-ready assets. It uses pluggable 3D backends (such as SAM3D, TRELLIS, and Hunyuan3D Pro) to generate meshes and URDFs. It can generate individual assets from text or images, create multi-room houses with controllable complexity, and compose complex layouts based on scene graphs parsed from natural language task descriptions.

Additionally, it features "3D Vibe Coding," which integrates with Claude Code to allow users to build and edit worlds through a natural-language dialogue interface using physics-validated skill calls.

Who it’s for

It is primarily for robotics researchers and AI developers working on embodied AI, reinforcement learning, and sim-to-real transfer, providing them with a scalable way to generate diverse training environments and assets.

Highlights

• Multi-Simulator Support: Consistent geometry and physics across SAPIEN, Isaac Sim, Isaac Gym, MuJoCo, Genesis, and PyBullet.
• Agentic World Building: Natural-language editing of 3D worlds via a Claude Code plugin.
• Sim-Ready Assets: Automatic generation of metric geometry, convex collision proxies, and VLM-inferred physical properties (mass, friction).
• Task-Driven Composition: Creates interactive 3D worlds including background, context objects, and robots based on task descriptions.
• Robot Learning Integration: Supports parallel gym environments for online training and evaluation of grasp quality.