AI agents wire your virtual commissioning model straight from CAD – and your vision AI is trained and tested in the running digital twin before the hardware exists.
Virtual commissioning AI means two things in practice. First, AI accelerates the engineering of the virtual commissioning model itself: AI agents such as Claude or Cursor connect to the digital twin through the free, open-source realvirtual MCP Server and create components, wire PLC signals and move drives – work that used to take days. Second, the digital twin commissions AI systems virtually: vision models are trained on synthetic data generated from the twin and tested against the simulated machine before a single real part has been produced.
realvirtual.io covers both directions on one Unity-based platform: the same model an AI agent helps build for virtual commissioning later generates the training data for your vision AI – and then runs the trained model live in the integration test with the real PLC.
The realvirtual MCP Server integrates AI agents directly into the engineering workflow: 150+ tools let the AI create components, define kinematics, wire PLCInput and PLCOutput signals and run drives in the simulation. The twin's knowledge is passed to the LLM in structured form – and because realvirtual Professional ships with full C# source code, the AI understands the framework from the very code you can read yourself.
The AI backend is yours to choose: EU-hosted or fully local. The result is a fully wired virtual commissioning model built from the CAD data in hours instead of days – reviewed by you, tested against your controller.
With realvirtual AIBuilder the digital twin becomes the training ground for industrial vision AI: it generates synthetic training data – images with pixel-perfect ground-truth labels for object detection, classification, segmentation and pose estimation – directly from your CAD geometry, with variators for lighting, position and texture.
Training a YOLO-based model is one click; the trained network is exported as a standard ONNX model and runs inside the simulation via Unity Sentis. Your camera system is virtually commissioned before the camera has been mounted – with no time-consuming collection of real image data on a running line.
AI in the twin is not a demo video – it is wired into the signal architecture. The vision model classifies, the result is written to PLC signals, the controller reacts, actuators respond: the complete AI-driven process is tested in the loop against the real controller – Software-in-the-Loop with PLCSim Advanced or Hardware-in-the-Loop with the real PLC over one of 25+ industrial interfaces.
That is the difference between an AI experiment and the virtual commissioning of an AI system: sequence logic, interlocks, timing and the AI model are validated together – at the desk, before on-site commissioning starts.
AI agents connect to the digital twin through the open-source realvirtual MCP Server and take over the repetitive part of model building: creating components, defining kinematics, wiring PLC signals, running test sequences. Engineers review and refine instead of wiring every signal by hand – a wired virtual commissioning model emerges from CAD data in hours instead of days.
Yes. realvirtual AIBuilder generates synthetic, automatically labeled training data from the digital twin, trains a YOLO-based model with one click and runs it as an ONNX model inside the simulation. The complete vision application – camera, AI model, PLC reaction – is tested virtually before any hardware exists.
Any MCP-capable agent – Claude and Cursor are the ones used most today. The realvirtual MCP Server is free and open source (MIT) with 150+ built-in tools, extensible with a C# attribute. It controls the Unity Editor and the running simulation: drives, sensors, PLC signals, robot kinematics.
The simulation runs locally on your engineering PC – CAD models and machine data stay in-house. For the AI agents you choose the backend yourself: an EU-hosted service or a fully local model. There is no forced cloud.
The MCP Server is free and open source. realvirtual Starter is free; realvirtual Professional costs €1,250 net, one-time per developer. The AIBuilder extension for synthetic data and vision AI training is €490 net, one-time. There are no runtime fees for delivered applications.
The trained network is exported as a standard ONNX model and executed inside the simulation via Unity Sentis. It classifies live rendered images and writes results to PLC signals, and the controller reacts – first Software-in-the-Loop, then Hardware-in-the-Loop against the real PLC. What passes this loop goes to the real machine.
Start free with realvirtual Starter and the open-source MCP Server – or go straight to Professional with 25+ interfaces and full source code.