Announcing SimReady Central: An Open Library of Simulation-Ready 3D Assets for Robotics and AI Factories

The Hugging Face Blog announced SimReady Central Library, an open collection of more than 2,000 simulation-ready (SimReady) OpenUSD assets hosted on Hugging Face. The library is free and ungated, and is positioned as a starting point for physical AI training environments and AI factory digital twins. Content spans infrastructure and equipment: robot arms and grippers, warehouse equipment, racks, and cooling infrastructure. Developers can inspect which capabilities each asset supports and combine assets from different suppliers into a single scene.
The premise is that a detailed 3D model, such as CAD, is only a starting point for simulation. SimReady assets carry the extra properties workflows depend on. For robot manipulation, that includes mass, collision geometry, and friction. For AI factory work, it includes equipment geometry plus the electrical or thermal metadata an application needs. OpenUSD is the open, extensible framework for describing and composing 3D scenes; using those scenes to train and evaluate autonomous machines also requires information about physical behavior, collisions, and sensor interactions.
The specification side is called SimReady Foundation. It provides specifications, profiles, and validation tools that make requirements explicit. A profile defines the logical grouping of features and capabilities an asset needs for a particular use case. The guidance is for developers to pick assets whose profiles match their workflow and compose them into an OpenUSD scene. In practice, this is the part that matters most for teams: an asset is only reusable across pipelines if everyone agrees on what properties it carries and how they are named, which is what a shared profile aims to fix.
Equipment manufacturers contribute through a submission skill that validates assets against the relevant SimReady profiles, packages them, and submits them for inclusion. Publishing gives OEMs discoverability with teams building robotics simulations and digital twins.
The announced contributors span several categories. For AI factory infrastructure, Vertiv is cited around racks, server racks, cooling equipment, and supporting infrastructure. Robotics OEMs named include Comau, FANUC, Franka Robotics, Robotiq, Standard Bots, and Universal Robots, contributing arms and grippers that can be combined with industrial equipment for warehouse and factory scenarios. Simulation asset provider Palatial contributes the NVIDIA Robotics 1K Dataset, released with NVIDIA Robotics, plus articulated assets generated from text, images, scans, datasheets, and CAD. LG Electronics contributes home-oriented assets such as refrigerators and dishwashers. ForgeCAD contributes articulated robotics props with high-quality textures for perception and manipulation training and synthetic data generation.
Search is central to how the library is meant to be used. NVIDIA USD Search powers natural-language search across participating datasets, and Hugging Face has integrated USD Search into its platform search with indexed USD properties and simulation metadata, so agents can find assets suited to a task. There is a search endpoint on the SimReady Central Hugging Face Space, and the GitHub repository documents index-to-library skills for teams that want to build their own asset browser.
The post walks through two illustrative workflows driven by an AI model. In the first, a prompt asks for a warehouse pick-and-place scenario in Isaac Sim using PhysX, finding a FANUC arm, a compatible Robotiq gripper, and warehouse assets, then generating Python to assemble the scene, configuring the task, and rendering a .mov preview with scripts and asset and runtime versions for reproducibility. The expected steps are querying the search API, referencing or payloading assets into an OpenUSD stage, generating assembly code, configuring the task, and rendering the preview. The second workflow has an independent software vendor build two AI factory layouts with racks and cooling equipment within supplied dimensions and aisle-clearance requirements, flagging overlaps and clearance violations, and saving each as an OpenUSD scene for comparison.
The post names OpenAI's GPT-6 Astra as a frontier model that can be paired with asset search and scene-authoring tools to find assets, write assembly code, and create scenario variations, and says OEMs can also use it to refine assets before validation and submission.
Suggested extension patterns are concrete: keep layout and task-specific changes in separate OpenUSD layers so one set of source assets supports multiple experiments; vary placements and equipment configurations; build application-specific tests in Isaac Lab for robotics or connect scenes to data and tools for power distribution or cooling analysis; and for factories, connect validated layouts to live BMS and IT telemetry to keep a digital twin synchronized with the facility.
The roadmap calls for more OEM partners and a growing asset library, with feedback sought on which assets and runtimes matter and where integration gaps appear. Bugs and requirements requests go to SimReady Foundation issues; OEM onboarding, an FAQ, CAD-to-SimReady skills, and a free OpenUSD learning path are linked as supporting resources. Outside contributors are pointed at Microduck, an open-source robot from Hugging Face and Pollen Robotics, as an example environment that SimReady assets can populate.
Why it matters: Teams building robot training or factory twins usually spend early effort on asset preparation rather than on the task itself, and a free, profile-validated catalog with metadata-aware search shifts that work onto manufacturers. The practical caveat, as an inference, is that usable coverage depends on which profiles your runtime supports and whether the specific equipment you need has been contributed; the breadth of named OEMs suggests momentum, but the library's value to any single team will track its own bill of materials. Expect asset search and scene authoring to become routine parts of simulation tooling, with reproducibility metadata treated as a first-class requirement.
Based on reporting from the original publisher. Visit the source for full context and later updates.