USDCraft: Geometry-Based Procedural Modeling of Articulated 3D Assets for Simulation
USDCraft converts text, images, or meshes into articulated Isaac Sim assets with doors, drawers, and hinges for robot training; artifacts are pending.
Key Takeaways
- USDCraft converts text, images, or meshes into simulation-ready articulated 3D assets with parts such as doors, drawers, and hinges.
- The system uses an agentic harness to automate modeling, reducing the difficulty and slow manual effort required for articulated objects.
- It targets Isaac Sim to provide movable, functional assets for robot training rather than static visual models.
- These assets help robot agents learn tasks like opening, closing, and handling articulated objects in simulation.
- USDCraft is currently pre-release: a project page is available, but code and dataset are pending and no public benchmark or timeline is given.
Meta Description: USDCraft uses an agentic harness to build simulation-ready articulated objects from text, images, or meshes, with doors, drawers, and hinges for robot training.
Tags: USDCraft, articulated assets, robot simulation, Isaac Sim, agentic harness
Featured Image: 
USDCraft is a system for turning real-world articulated objects into simulation-ready assets for robot training. It uses an agentic harness to take three inputs—text, images, or meshes—and reconstruct objects with doors, drawers, and hinges. The target is Isaac Sim, which matters because robot agents need moveable parts, not static shapes, to learn opening, closing, and handling tasks.
Background: Why Articulated Objects Are Hard in Simulation
Articulated objects are difficult to simulate because each part needs joints, motion ranges, and collision behavior. Manual modeling is slow and limits the variety of environments available for training.
Development: From Three Inputs to Isaac Sim Assets
USDCraft addresses this by using an agent workflow to convert three input types into usable assets. The system can accept text, images, or meshes, then build articulated structures such as doors, drawers, and hinges. This creates assets that can be used in simulation instead of only visualized.
Latest: Project Page Available, Code and Dataset Pending
A project page is listed, while two open artifacts—code and dataset—are pending release. No public benchmark or timeline is provided.
Current Stage: Pre-Release Tool for Robot Training Assets
USDCraft is currently pre-release: its reconstruction method and Isaac Sim target are described, but open artifacts have not yet been published.
Comments
No comments yet. Be the first to comment.