Hugging Face now hosts 2,000+ free simulation-ready 3D assets for robots and AI factories

Nvidia has opened SimReady Central Library, a collection of more than 2,000 simulation-ready OpenUSD assets hosted on Hugging Face, according to an official post published October 21. Nvidia describes it as a free, ungated starting point for robotics training environments and AI factory digital twins.

What is in the library and who supplies it

A SimReady asset is a 3D model that already carries the properties a simulator needs, such as collision geometry, mass and friction. Developers normally add these by hand to a CAD model. According to the post, the library covers robot arms and grippers, warehouse equipment, racks and cooling infrastructure, and developers can check which capabilities each asset supports before combining parts from different suppliers into one scene.

The post names the contributors by sector:

• AI factory equipment: Vertiv, for racks, cooling and supporting infrastructure.
• Robot arms and grippers: Comau, FANUC, Franka Robotics, Robotiq, Standard Bots and Universal Robots.
• Home robotics: LG Electronics, which plans to share assets such as refrigerators and dishwashers.
• Asset providers: Palatial, with the NVIDIA Robotics 1K Dataset and articulated assets generated from text, images, scans, datasheets and CAD, and ForgeCAD, with articulated props.

What qualifies as SimReady is set by SimReady Foundation, which publishes specifications, profiles and validation tools. A profile bundles the features an asset needs for one use case, for example joint limits for articulated machines, semantic labels for synthetic data, or sensor attributes for rendering. Manufacturers validate their models against the relevant profile, package them and submit them through a SimReady Central submission skill. Nvidia says its roadmap includes more OEM partners and a larger library.

Natural-language search aimed at people and agents

Discovery runs through NVIDIA USD Search, which powers natural-language queries across participating datasets. Hugging Face has also integrated USD Search into its platform search, with indexed USD properties and simulation metadata, so that agents can find assets suited to a given task.

The post leans heavily on AI-assisted authoring. It describes two sample prompts for OpenAI’s GPT-6 Astra: one to assemble a warehouse pick-and-place scene in Isaac Sim with PhysX using a FANUC arm and a Robotiq gripper, and one to build two alternative AI factory layouts and check them for equipment overlaps and aisle-clearance violations. Nvidia says it expects the model to query the search API, reference assets into an OpenUSD stage, write the assembly code and render a preview. These are described as expected workflows, and the post publishes no results or timings from them.

Free assets shift the cost to validation

The economics are straightforward. Preparing CAD for physics is repetitive engineering work, and a shared library moves that cost from every developer to the manufacturer that makes the equipment. Developers get free parts; manufacturers get distribution among the teams that simulate their products before buying them. The post says OEMs manage their own Hugging Face datasets, so the library is a federation of datasets rather than a single Nvidia-owned store.

Two details outside the announcement add context. Nvidia’s documentation describes SimReady Foundation as an open specification layer on top of OpenUSD, shipping validation rules and feature adapters that convert an asset between profiles, and organized into features, profiles and tiers. The same documentation exists in several releases, including 2026.07.1 and 2026.08.0, and example profile names differ between them. In practice, an asset is only as reusable as the profile version it was checked against, so teams will need to confirm that version before mixing suppliers.

The post also leaves open how permissive the licensing is. It says the library is free and ungated but names no license in the text we reviewed, and individual OEM datasets may carry their own terms. Nvidia made a similar choice with its Nemotron 3.5 retrieval models, where, as we reported, the announcement pointed readers to model cards for license details. The same check applies here: read the dataset card before building a commercial product on any asset.

What the announcement leaves unproven

All claims come from Nvidia. The post does not say how many of the 2,000 assets belong to each category, how many meet which profile, or how well the assets behave across runtimes beyond Isaac Sim and Isaac Lab. It asks the community to report which runtimes they use and where integration gaps appear, which suggests coverage is still uneven. Developers can start with the library on Hugging Face, the OpenUSD learning path and the Isaac Sim and Isaac Lab tools that the post links. For Nvidia, which also recently pushed AI factories as a product line with its open-weight model releases, a common asset format keeps simulation work tied to OpenUSD and its tooling.