NVIDIA Outlines 5 Steps for Preparing SimReady Assets for Robots with AI Agents

NVIDIA has released a five-step guide on its Developer Blog for turning CAD models into SimReady assets for robot simulation. The workflow builds on Omniverse libraries, the SimReady Foundation specifications and agentic skills, which bring frontier models into the process. Anyone who has built a scene for a robot knows the problem: converting geometry to OpenUSD covers only a small part of the job.

Why converting to OpenUSD is not enough

A CAD file describes the shape of a part. A simulator needs far more. Materials determine how a surface looks and how it interacts with other bodies. Collision geometry defines a simplified outline that the physics engine uses to compute collisions. Joints describe which links move, and how. Mass, friction and other physical parameters decide whether an object behaves like the real thing.

If any of these is missing or configured incorrectly, the simulation keeps running, but the results are useless. A manipulator falls through the table. A part slides as if on ice. A joint rotates 360 degrees where the real hardware has a stop. NVIDIA states the requirement plainly: properties must be configured and validated before you start testing the robot's behavior.

What the workflow consists of

NVIDIA builds the process from three components:

• Omniverse libraries handle asset conversion and validation.
• SimReady Foundation specifications define which properties an asset must have and how to check that it is ready for simulation.
• NVIDIA agentic skills let a frontier model walk an asset through the process step by step instead of configuring it by hand.
• The result is described as a structured conversion-and-validation scenario split into five steps.

The available excerpt of the post does not say which models work with the skills. It does not give any time-savings metrics either. The specific content of the five steps appears only in the full version of the post.

Where the bottleneck is

Asset preparation remains one of the most tedious parts of robotics development. Industrial companies have huge CAD libraries, and every model needs its own round of physics fiddling. When a standard defines what a "ready" asset is and validation can run automatically, the process becomes reproducible. That is the main value of SimReady as a specification: it turns a vague "seems to work" into a verifiable list of requirements.

The agentic layer adds speed. A model can read the assembly structure, propose collision shapes, place joints and run the validator. But there is a caveat here, and this is my assessment, not NVIDIA's claim. A specification validator checks the completeness and consistency of the data. It does not know that the real friction of rubber on aluminum differs from the default value. Parameters that affect sim-to-real transfer still require measurements, or at least expert review.

What teams should do

If you already have a CAD library and work with Isaac Sim or another OpenUSD-based stack, start with a single asset of medium complexity. Run it through all five steps of the guide and compare the model's behavior in simulation with that of the physical sample: mass, tipping moment, joint travel.

Then divide responsibility. Give the agent the routine work: structure, naming, basic collisions, running the checks. Leave joint limits, mass, friction and anything that affects safety to the engineer. Only once the gap from reality becomes predictable does it make sense to scale the process to hundreds of parts. The full text of the guide is on the NVIDIA Developer Blog, and that is where to look for the details of each of the five steps.