EmbodiedGenV2
An Agentic, Simulation-Ready 3D World Engine
for Embodied AI
From intent to executable 3D worlds
1Horizon Robotics · 2WuwenAI
Paper
Code
Video
HuggingFaceGenerative simulation infrastructure for embodied AI
From text, images, and natural-language dialogue to simulation-ready 3D worlds — assets, interactive scenes, and large-scale environments, deployable across every major simulator.
Generate Sim-Ready Assets
Metric geometry, collision proxies, and physical properties in one package.
↓Scale Large-Scale Scenes
Multi-room, navigable environments with editable furniture instances.
↓Compose Task-Driven Worlds
Natural-language tasks become physically stable, simulator-ready layouts.
↓Edit Vibe Coding
Refine a deployable world state through natural-language dialogue.
↓Export Cross-Simulator
One standardized layout, six engines, zero manual adaptation.
↓Train Closed-Loop Learning
Generated environments as online RL worlds for robot policies.
↓View the system overview figure
Sim-Ready 3D Assets
Real generated assets — drag to rotate. Each carries metric geometry, a collision proxy, physical properties, and cross-simulator interfaces.
Pipeline text or a single image → a sim-ready asset
- 01
Input Preparation
Text → image, or segment the foreground — even under occlusion.
- 02
3D Generation & QC
Pluggable image-to-3D, gated by semantic, geometric & aesthetic checks.
↻ auto-retry - 03
Geometry & Texture
Mesh repair, simplification, convex decomposition, baked texture.
- 04
Physical Recovery
A VLM infers real-world scale, mass and friction for metric rescaling.
- 05
Cross-Format Export
One URDF intermediate → URDF · USD · MJCF for every major engine.
View the full pipeline figure
Beyond rigid bodies — soft-body simulation
The same generate-and-export path reaches soft bodies: 12 text-conditioned garments deploy as deformable meshes in Genesis, with no manual preparation.
12 generated garments · per-vertex displacement · Genesis
Large-Scale Scenes
Beyond tabletops — multi-room, navigable, instance-editable houses, generated as sim-ready backgrounds at a controllable complexity tier.
View the large-scale generation pipeline
Task-Driven Worlds
From a natural-language task, EmbodiedGen parses a Scene Graph and composes a physically stable, directly loadable layout — then settles it under gravity in simulation.
View the scene-generation pipeline
Vibe Coding for Sim-Ready 3D Worlds
Build and edit worlds through natural-language dialogue. Each instruction is a bounded, physics-validated skill call that preserves a deployable, sim-ready world state — add, remove, replace, refine.
One World, Every Simulator
One standardized layout — no manual adaptation. The same generated scene loads with consistent geometry, collision, textures, and physical metadata across six mainstream physics simulators.
Closed-Loop Robot Learning
Generated worlds are not just viewable — they are online training environments. Policies trained purely in EmbodiedGen-generated worlds transfer to real robots.
Results from a companion sim-to-real RL study; policies trained in EmbodiedGen-generated worlds.
View the policy-learning & deployment figure
BibTeX
@misc{wang2026embodiedgenv2agenticsimulationready,
title={EmbodiedGen V2: An Agentic, Simulation-Ready 3D World Engine for Embodied AI},
author={Xinjie Wang and Liu Liu and Taojun Ding and Andrew Choi and Chaodong Huang and Mengao Zhao and Ziang Li and Jackson Jiang and Chunlei Yu and Shengxiang Liu and Wei Xu and Zhizhong Su},
year={2026},
eprint={2607.07459},
archivePrefix={arXiv},
primaryClass={cs.RO},
url={https://arxiv.org/abs/2607.07459},
}
@misc{wang2025embodiedgengenerative3dworld,
title = {EmbodiedGen: Towards a Generative 3D World Engine for
Embodied Intelligence},
author = {Xinjie Wang and Liu Liu and Yu Cao and Ruiqi Wu and
Wenkang Qin and Dehui Wang and Wei Sui and Zhizhong Su},
year = {2025},
eprint = {2506.10600},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2506.10600}
}