v2.0.0

EmbodiedGen v2.0.0 Release Notes

New Features

EmbodiedGen V2: Agentic, Simulation-Ready 3D World Engine

This major release upgrades EmbodiedGen from an asset-generation toolkit into an agentic 3D world engine for embodied AI. V2 supports generating, editing, composing, exporting, and validating simulation-ready assets, rooms, and task-driven interactive worlds.

3D Vibe Coding

Added a Claude Code plugin with /embodiedgen:* slash commands for building and editing 3D worlds through natural-language dialogue. Users can generate assets, create indoor scenes, compose layouts, run simulation, convert assets, and process scene objects through bounded, physics-aware skill calls.

Large-Scale Indoor Scene Generation

Added room and multi-room house generation with controllable complexity tiers, navigable layouts, instance-editable objects, bird’s-eye rendering, USD export, and trajectory generation.

Pluggable 3D Generation Backends

Expanded the image-to-3D stack with pluggable backend support, including SAM3D, TRELLIS, and Hunyuan3D Pro via a unified model-selection interface.

Part-Level Affordance Annotation

Added affordance annotation tools for generated assets, including functional part segmentation, part-level semantics, and simulation-validated 6-DoF grasp pose generation.

Soft-Body Simulation

Added deformable object support for text-conditioned garment generation and Genesis-based cloth simulation, including conversion, simulation, and displacement heatmap rendering utilities.

Robot Learning Utilities

Added workflows for parallel simulation environments and grasp-quality evaluation, supporting closed-loop robot learning experiments with EmbodiedGen-generated worlds.

Articulated Object Loading

Added support for loading articulated objects in simulation workflows.
(PR #65)

Platform & Simulator Support

One World, Every Simulator

Improved the standardized asset and layout pipeline for deployment across SAPIEN, Isaac Sim, Isaac Gym, MuJoCo, Genesis, and PyBullet with consistent geometry, collision, textures, and physical metadata.

AMD ROCm Support

Added AMD ROCm gfx942 support for the image-to-3D generation stack, including a dedicated Docker setup, one-shot installation script, ROCm compatibility shims, and import smoke checks.
(PR #72)

Fixes

• Fixed missing inpainting model by switching from stabilityai/stable-diffusion-2-inpainting to sd2-community/stable-diffusion-2-inpainting.
  (PR #69)

Documentation Improvements

• Updated README for EmbodiedGen V2.
• Added the new project landing page with videos, model previews, posters, and Open Graph assets.
• Reorganized documentation under the new docs structure.
• Added or expanded tutorials for room generation, 3D vibe coding, affordance annotation, soft-body simulation, robot learning, digital twin creation, image/text-to-3D, texture editing, layout generation, scene generation, and simulator export.
• Added installation scripts for submodules, basic setup, CUDA 12.6, room generation, scene generation, affordance tools, and the Claude Code plugin.
• Added affordance demo assets and download guidance.

Summary

EmbodiedGen v2.0.0 is the largest release so far. It introduces an end-to-end workflow for generating simulation-ready 3D assets, scaling them into large indoor scenes, composing task-driven interactive worlds, editing them through natural-language agent commands, and deploying them across major robotics simulators.

This release also improves hardware compatibility, especially with AMD ROCm support, and significantly expands documentation, demos, tutorials, and installation paths for researchers and developers building embodied AI environments.

What's Changed

• feat(sim): Add support for articulated object loading. by @HochCC in #65
• chore(docs): Update feedback wanted. by @HochCC in #68
• fix(model): Fix inpainting model missing. by @HochCC in #69
• Add AMD ROCm (gfx942) support for the image→3D generation stack by @ZJLi2013 in #72
• feat(all): Big update for V2 version. by @HochCC in #73
• chore(docs): Update for the webpage. by @HochCC in #74
• chore(docs): Update README and add affordance download. by @HochCC in #75
• chore(docs): Update landing page video. by @HochCC in #76
• chore(docs): Update video link, landing page. by @HochCC in #77
• chore(docs): Update video link. by @HochCC in #78
• chore(docs): Update landing page format error. by @HochCC in #79
• fix(env): Fix env and installation issue. by @HochCC in #80
• chore(docs): Update landing page url. by @HochCC in #81

Full Changelog: v0.1.7...v2.0.0

v0.1.7

🆕 EmbodiedGen v0.1.7 Release Notes

🚀 New Features

Refined Texture Back-Projection Pipeline

Improved the texture back-projection process with more accurate projection and fusion, enhancing texture quality and visual consistency on generated 3D assets. by @HochCC in PR #58

SAM3D Integration for Image-to-3D Modeling

Integrated SAM3D as a new image-to-3D model option, expanding EmbodiedGen’s supported reconstruction methods and improving flexibility in asset generation workflows. by @HochCC in PR #61

🧩 Summary

This release focuses on enhancing the visual quality and modeling flexibility of EmbodiedGen.
With refined texture back-projection and the addition of SAM3D as an image-to-3D option, users can generate higher-quality assets and choose from a richer set of reconstruction pipelines for embodied AI and simulation scenarios.

Full Changelog: v0.1.6...v0.1.7

v0.1.6

🆕 EmbodiedGen v0.1.6 Release Notes

🚀 New Features

• Infinigen Asset Conversion Support
  Added a dedicated script for converting Infinigen assets into EmbodiedGen-compatible format.
  (PR #47)

• Cross-Framework Asset Conversion Interface
  Introduced cvt_embodiedgen_asset_to_anysim interface for seamless conversion of EmbodiedGen assets to AnySim format.
  (PR #48)

🧱 Documentation Improvements

• Added MkDocs integration to the project documentation management system.
  (PR #50)
• Fixed MkDocs build workflow in GitHub Actions.
  (PR #51)
• Updated and refined README and general documentation content.
  (PRs #49, #52–55)
• Enhanced docstrings across codebase and synchronized improvements into documentation.
  (PR #56)

🧩 Summary

This release focuses on improving asset conversion capabilities and documentation quality.
Developers can now easily convert assets between EmbodiedGen, Infinigen, and AnySim, with better-maintained docs and build pipelines.

Full Changelog: v0.1.5...v0.1.6

v0.1.5

• Support one-click import into all mainstream simulators for assets generated by EmbodiedGen.

Simulator	Conversion Class
isaacsim	MeshtoUSDConverter
mujoco	MeshtoMJCFConverter
genesis / sapien / isaacgym / pybullet	EmbodiedGen generated .urdf can be used directly

• Optimize the storage volume of EmbodiedGen generated assets.

What's Changed

• feat(sim): Support all simulators and opt asset storage size. by @HochCC in #40
• fix(layout): Fix layout axis order after urdf origin rotation change. by @HochCC in #41
• feat(layout): Expose limit_reach_range and max_orient_diff in bfs_placement. by @HochCC in #42
• chore(layout): Misc update for layout sim. by @HochCC in #43
• chore(docs): Update README. by @HochCC in #44
• fix(backproject): Fix mesh resimply error by adding mesh.triangulate(). by @HochCC in #45
• fix(sim): Fix .obj mesh compose to solve collision issue. by @HochCC in #46

Full Changelog: v0.1.4...v0.1.5

v0.1.4

• Supports importing the output of the EmbodiedGen Layout Generation module with a single gym.make command to create multiple parallel simulation environments and interactive 3D worlds.
• Supports coacd-based mesh decomposition for better efficiency and accuracy in physical simulations.
• Optimize backprojected texture quality and add texture-cli.

What's Changed

• feat(sim): Optimize support for downstream simulators and gym. by @HochCC in #36
• feat(sim): Add auto scale in convex decomposition. by @HochCC in #37
• feat(texture): Optimize backprojected texture quality and add texture-cli. by @HochCC in #38

Full Changelog: v0.1.3...v0.1.4

v0.1.3

Feature

Release layout-cli 🏞️ Layout(Interactive 3D Worlds) Generation, generate interactive 3D scene from task description.

v0.1.2

Feature

🖼️ Single Image to Physics Realistic 3D Asset

• Generates watertight, simulation-ready 3D meshes with physical attributes.
• Auto-labels semantic and quality tags (geometry, texture, foreground, etc.).
• Produces 2K textures with highlight removal and multi-view fusion.

📝 Text-to-3D Asset Generation

• Creates realistic 3D assets from natural language (English & Chinese).
• Filters assets via QA tags to ensure visual and geometric quality.

🎨 Texture Generation & Editing

• Generates 2K textures from mesh and text with semantic alignment.
• Plug-and-play modules adapt text-to-image models for 3D textures.
• Supports realistic and stylized texture outputs, including text textures.

🌍 3D Scene Generation

• Generate a color mesh and 3DGS per scene.

Others:

• provide Dockerfile.