These five steps explain how to create simulation-ready assets for robotics using frontier artificial intelligence models.

Meta description: A SimReady workflow converts ABB Robotics YuMi STEP files into validated OpenUSD assets for NVIDIA Isaac Sim using Omniverse libraries and GPT-6 Astra.

Tags: SimReady, NVIDIA Isaac Sim, OpenUSD, ABB YuMi, Omniverse AI Skills

Featured image: Abstract friendly illustration of a dual-arm robot being converted from CAD geometry into a simulation-ready asset, with checkmarks and cubes; no real people, no logos, no text.

The SimReady workflow helps developers prepare an ABB Robotics YuMi robot for NVIDIA Isaac Sim by converting STEP files into OpenUSD, checking appearance, configuring physics, validating against SimReady Foundation requirements, and running a dual-arm pick-and-place test. NVIDIA Omniverse libraries, guided by SimReady Foundation specifications and agentic NVIDIA skills, provide the structure. GPT-6 Astra is shown as one frontier AI model example that can write Python code to call those tools, reducing manual asset preparation.

Why SimReady matters before testing robot behavior

SimReady defines requirements for preparing OpenUSD assets for specific simulation use cases. Developers select a SimReady profile for their intended use and validate the asset against its requirements. Profile-specific checks validate structure, materials, and physics properties, then flag issues for developers to fix and recheck. A robot can look correct but fail in simulation if it lacks simulation properties. Missing collision geometry can allow objects to pass through a gripper, incorrect joints can prevent coordinated motion, and mass and friction directly affect whether a grasp remains stable.

The workflow uses OpenUSD to represent geometry, assembly structure, materials, and simulation properties in a common asset representation. SimReady Foundation provides specifications and validation guidance. NVIDIA Omniverse libraries provide capabilities for converting, inspecting, and preparing OpenUSD content. NVIDIA Isaac Sim provides the environment for configuring physics and testing robot behavior. NVIDIA Omniverse AI Agent skills, such as the CAD-to-SimReady skill, guide agents through asset conversion, material and physics assignment, and validation.

Setup for the YuMi example

The example uses an ABB YuMi robot with both arms and grippers. Cubes and a Sharpie marker were used as demonstrations in the final step. Camera-based perception can be added in a separate workflow. GPT-6 Astra is used here as one example of a frontier AI model that can help write Python code to call NVIDIA Omniverse libraries at each stage, but the workflow can be applied across many models and robot systems.

Before starting, install Isaac Sim and create a folder inside the Isaac Sim directory named reference. Add robot specifications, STEP files, and reference images and videos to that folder. Install the frontier AI model; this example used GPT-6 Astra. Launch Codex CLI in the Isaac Sim directory. Launching Codex inside that directory is important because the agent needs to find the workflow skills. The example used Isaac Sim 6.1 and Codex CLI, with Isaac Sim launched in windowed mode. Use an example prompt in the active Codex CLI to start Isaac Sim. The remote Python server in Isaac Sim lets the agent talk through Python to the live Isaac Sim application, so confirm the connection before continuing. At each step, adapt the sample prompt to the task, then review simulation results and validation feedback to identify issues and refine the asset.

Five-step SimReady workflow

1. Import the robot from STEP files. Provide the ABB Robotics YuMi robot STEP files, the linked ABB Robotics video and image, the SimReady Foundation specification repository, and the CAD-to-SimReady skill in the reference folder. Isaac Sim 6.1 is packaged with simready-foundation-tier-core, which supplies robot requirements, related skills, and validation guidance. Ask the agent to use the inputs to guide conversion to OpenUSD and import into Isaac Sim. The skills used are omniverse-cad-to-simready, omniverse-cad-to-usd, and isaac-sim-remote. Before continuing, confirm the robot asset is imported and available for inspection, the task and supplied requirements are understood, and missing information is identified.

1. Match the visual appearance. Inspect the imported robot to confirm expected materials and textures are present and loaded correctly. Compare the asset with the supplied ABB images and videos to check scale, orientation, and part placement. Ask the agent to add or refine materials and textures using those references. The skills used are isaac-sim-remote and usd-pipeline. The agent visually matches materials by adjusting colors, metallic response, and roughness while preserving the CAD geometry and mounting transforms. These material settings were visual estimates, without measured material reconstruction or AI-generated textures. Before continuing, confirm the robot has expected scale, orientation, parts, and appearance, and document any missing materials, textures, or visual discrepancies.

1. Configure physics properties. At this stage, the OpenUSD asset contains visual geometry, but joints, rigid bodies, and collision geometry have not yet been configured. Ask the agent to configure and validate physics properties for both arms and grippers. The datasheet in the reference folder helps the agent make sure joint ranges and speed limits match the simulated robot. If manufacturer data is incomplete, the agent can identify missing parameters and check technical manuals or available URDF robot descriptions for the exact model. Where mass properties remain unavailable, it can estimate them from geometry using documented density or mass-distribution assumptions, following NVIDIA mass-property guidance. In the example, when context was insufficient, the agent pulled a public URDF for the robot and used it to identify the correct axes for joints and the zero configuration. The initial pose may vary.

The simulation assumptions were: joint axes and zero configuration came from a public URDF; per-link masses were estimated from STEP geometry volumes and normalized to manufacturer-specified totals of 38 kg for the robot and 0.28 kg for each gripper; centers of mass and inertia assumed solid parts of uniform density, without individually modeling internal motors, gearboxes, or wiring; contact properties assumed static friction of 0.8, dynamic friction of 0.6, and zero restitution; collision geometry and appearance used approximate convex collision shapes, surface roughness, and material response; and the parameters were not calibrated against physical YuMi measurements. The skills used are usd-articulation, physics-simulation, isaac-sim-validator, and isaac-sim-remote. Before continuing, confirm joints move around expected axes and within limits, base mounting and gripper motion behave as configured, and estimated properties and validation failures are documented.

1. Validate the SimReady properties. Ask the agent to run the selected SimReady Foundation checks, review findings, and revise the asset as needed. The checks covered units, asset dependencies, geometry, materials, rigid bodies, joints, drives, and articulation. Isaac Sim asset checks verified scale, orientation, part placement, gripper mounting, material coverage, and resolved texture dependencies through the validation engine inside Isaac Sim. All 21 rigid bodies had positive masses and physically admissible inertia tensors, and runtime values matched authored values. The joint review checked axes, frames, limits, drives, finger mimic behavior, and fixed-base mounting. Runtime behavior tested arm and gripper motion, contact, release, and collisions along the tested trajectories. These checks validated the simulation implementation, but did not establish agreement with real-world robot dynamics or collision safety across every possible pose. The skills used are usd-articulation, physics-simulation, isaac-sim-validator, and isaac-sim-remote. Before continuing, confirm the sequence meets agreed acceptance criteria and selected SimReady checks pass, and document tuning changes, estimated properties, and unresolved failures.

1. Run the final task. Perform a final validation in Isaac Sim combining visual, physics, and SimReady checks with a pick-and-place task. In the cube example, ask the agent to add color targets and a box to the test scene, define target locations and color metadata, and select matching color targets using scene metadata. The simulation used dimensionless contact-friction coefficients of 0.8 static and 0.6 dynamic; the grasp tests passed with these settings, but the coefficients were not measured or calibrated against physical finger-object contact. The test cubes were 45 mm and 40 g, and the step 3 assumptions applied. Both arms and grippers completed four pick-and-place cycles in total during a 122.2-second physics simulation. Each arm grasped its cube, lifted it, held it for two seconds, transferred it, and released it onto its matching color target. Each arm then picked up its cube again and placed it inside the box. Finger contact and friction carried the cubes, without attachment joints, kinematic holds, or direct cube-pose updates during simulation. All cycles passed the defined grasp, hold, transfer, release, and placement criteria. The MP4 and GIF were generated from this simulation. A second demonstration used a Sharpie marker; the source also begins another pick-and-place example, but its details are not supplied.

Common mistakes

Common mistakes include treating visual import as enough, skipping collision geometry, relying on uncalibrated mass and friction assumptions, and assuming SimReady checks prove real-world dynamics. Developers should adapt the SimReady profile to the intended task, run the checks, and record any estimated properties or unresolved failures before moving to camera-based perception or wider robot behavior tests.