How Multi-Agent Orchestration Between Onshape and NVIDIA Improves CAD-to-Simulation Workflows

As industries strive to deploy autonomous robotics, intelligent factories, and sophisticated manufacturing systems, engineering teams require a new generation of toolsets. These tools must transcend traditional documentation meant strictly for efficient and reliable manufacturing. They must capture and communicate semantic design intent downstream to facilitate the simulation-driven workflows needed to design today’s advanced products.

A new workflow showcases a deeper connection between PTC’s Onshape and NVIDIA Isaac Sim. Using an AI agent pipeline, the workflow incorporates automatic preparation of CAD models for simulation by verifying geometry, inferring mates, correcting issues, and delivering a simulation-ready asset without manual intervention.

Modeling the Physical World and Physical AI

To build AI that can perceive, reason, and act in the physical world, often referred to as Physical AI, teams must first build highly accurate virtual models of that world. Physical AI can’t be trained purely in text or 2D images; it requires an accurate understanding of real-world physics like mass, friction, gravity, force, and material properties.

Modeling the physical world for AI training demands three key components:

1. Geometric Fidelity: The exact 3D representation of shapes and assemblies.
2. Physical Fidelity: Realistic dynamics, mass distributions, and constraints (e.g., joints, hinges, gears).
3. Semantic Fidelity: Understanding what an object is and how it functions (e.g., knowing that a specific cylinder is a wheel meant to rotate around an axle, rather than just a collection of faces).

Historically, CAD software was optimized strictly for geometric fidelity and manufacturing documentation.

CAD platforms such as Onshape by PTC are elevating workflows beyond mere geometric modeling. The future of CAD lies in becoming the single source of truth, not just for how a product looks or is manufactured, but how it behaves in the real world.

Core Technologies for Physically Accurate Simulation

At the core of this stack are two foundational technologies: OpenUSD and the open NVIDIA Isaac Sim framework.

OpenUSD

OpenUSD (Universal Scene Description) is far more than just another 3D file format; it is an extensible, open, hierarchical scene description framework developed initially by Pixar. It allows teams to work on the same 3D scene simultaneously, supporting:

• Layering and Composition: Modifying a scene without altering the original asset files.
• Rich Physics Schemas: Out-of-the-box definitions for rigid bodies, joint limits, colliders, and materials.
• Lossless Data Interchange: Retaining complex structural and physical properties across different software suites.

NVIDIA Isaac Sim

Designed specifically for robotics, the open NVIDIA Isaac platform leverages Omniverse libraries to train and test autonomous machines: Isaac Sim is a robotics simulation framework that provides photo-realistic, physically accurate virtual environments. It allows developers to test physical interactions in real time.

AI Agents in Onshape

Onshape is developing AI agents that can be added to projects like any other teammate, with defined permissions that can be revoked at any time. Their actions are fully tracked, visible, and reversible, maintaining the same level of control teams expect from human collaborators.

Agents can analyze designs, provide feedback, answer questions about model data, export deliverables, fix failing features, update parameters, modify geometry based on defined goals, automate repetitive tasks, and assist with drawing, creating a true force multiplier rather than a passive assistant.

Utilizing the Onshape MCP (Model Context Protocol) server, a toolkit for custom agents, will give designers and engineering teams a stable, open interface to connect their own strategies directly to Onshape. Teams can build agents trained on company-specific data and tuned to their own needs, extending Onshape’s already unbounded integration abilities.

Multi-Agent Orchestration for Industrial AI Use Cases

To rapidly execute workflows from Onshape to the Isaac Sim environment, a multi-agent orchestration framework is deployed. This system uses specialized AI agents to automate the translation and enrichment of data:

Onshape AI Agent

Role: Acting as the orchestrator, routes specialized Onshape agents to NVIDIA libraries and skills as necessary.

Action: Audits the readiness of the model in context for Isaac Sim simulations. Runs SimReady Publisher to create a USD file. Identifies missing mates, based on inputs from the Content agent’s physics agent. Identifies missing materials based on the inputs from the Content agent’s Material agent. Validates if the translated model is ready for Isaac Sim simulation by running checks for consistent unit systems.

The Material Agent

Role: Automates the assignment of physically plausible, sensor-ready materials (PBR/MaterialX/OpenPBR) to 3D meshes using visual VLM reasoning.

Action: Generates multi-view renders of the incoming CAD meshes, feeds these images to a VLM, and cross-references a designated material library (like NVIDIA’s Physical AI material dataset) using fuzzy validation. It then writes the correct material binding APIs directly back to the target USD prims. Additionally, it can ingest engineering specification PDFs via RAG to resolve highly specific material requirements.

The Physics Agent

Role: Classifies components and resolves structural physical properties (friction coefficients, mass distribution, collision geometry) based on physical-world reasoning.

Action: Runs VLM-based visual analysis on the component structure to determine asset types (e.g., recognizing rotating drum/rotor mounted on a base). It infers friction, density, and restitution parameters, setting up collision bounds and serializing this structured data back into the USD file under Physics Schemas.

From Static CAD Model to Physics-Rich Simulation

The journey from a static CAD model to a dynamic, physics-rich simulation environment has historically been manual, tedious, and prone to data loss.

By combining Onshape’s cloud-native architecture, OpenUSD, NVIDIA Isaac, and the power of multi-agent orchestration, we are fundamentally redefining this workflow.