Industrial

The Engine Room of AI

For Physical AI, we need data volumes that cannot be handled manually. In the real-world lab, we have therefore developed our own highly automated pipeline tested. It is based on open-source technology (Blender) and functions as a “factory” that transforms CAD models into intelligent training data. 

Step 1: Physics as a Foundation

Physical AI must be physically plausible. That’s why our pipeline doesn’t start with graphics, but with physics. 

• Sanity Check: An integrated simulation drops components virtually. Objects that are stuck together or are suspended in mid-air are eliminated. Only what is physically possible may be part of the training. 

• Stress Test: We deliberately add interference objects (“distractors”) and geometric anomalies to prepare the AI for the chaos of the real world. 

Step 2: Controlled Chaos

To ensure robustness, we systematically vary the simulation parameters: 

• Lighting: Dynamic lighting scenarios simulate day, night, and artificial indoor lighting. 

• Context: "Background Swapping" places industrial parts in front of irrelevant backgrounds so that the AI learns to maintain focus. 

• Material: Wir simulieren blitzschnell verschiedene Oberflächenbeschaffenheiten auf demselben Objekt. 

Step 3: Industry Standard

The output from our pipeline is ready to use. We generate standardized formats (HDF5, COCO, BOP) that can be seamlessly integrated into modern AI training processes. The result is datasets that contain not only images but also depth information and segmentation masks.