NVIDIA Uses GPT-6 Astra to Build SimReady Robot Assets

NVIDIA has introduced a structured five-step workflow that leverages frontier AI models, such as GPT-6 Astra, to convert CAD files into simulation-ready OpenUSD assets. Using the SimReady Foundation specifications and agentic Omniverse skills, this process dramatically reduces the manual effort required to prepare virtual robots for testing. The system was demonstrated using an ABB YuMi robot imported from STEP files, showcasing how AI can write Python code to call Omniverse tools at each stage of development.

During the demonstration, the GPT-6 Astra model assisted in configuring the robot's physics properties using manufacturer datasheets and a public URDF. The setup defined joint axes, limits, and mass properties, including a 38 kg mass for the robot and 0.28 kg for each gripper. The simulation assumed contact friction coefficients of 0.8 static and 0.6 dynamic, with zero restitution. The AI calculated the centers of mass and inertia for 21 rigid bodies, assuming uniform density for the solid parts.

To ensure the asset met strict SimReady requirements, developers ran validation checks inside Isaac Sim 6.1 using the Codex CLI. The validation engine verified units, geometry, materials, rigid bodies, joints, drives, and runtime behavior. In a final test, the virtual YuMi robot successfully completed four pick-and-place cycles during a 122.2-second physics simulation. The robot manipulated 45 mm, 40 g test cubes and a 140 mm, 10 g Sharpie marker modeled from a reference image, executing grasps and releases without relying on artificial attachment joints.

For robotics developers, this workflow eliminates the tedious manual labor of translating raw CAD geometry into functional simulation assets. By utilizing agentic skills like the CAD-to-SimReady skill, practitioners can quickly bridge the gap between design and testing. This ensures that virtual robots behave realistically in environments like Isaac Sim before any physical deployment, accelerating development timelines and reducing errors in physical AI training.