Industrial

From Hypothesis to Standard

Through thousands of experiments ("ablation studies"), we have analyzed which factors make a synthetic dataset truly effective. These findings serve as a guide for the development of Physical AI. 

The Formula for Success: What Really Matters? 

Data show a clear hierarchy of importance for robust perceptual systems: 

1. Light & Shadow (+2–4%): Models that learn only under ideal conditions fail immediately in real-world scenarios. Variation in lighting is essential. 
2. Physics (+1–2%): A physically accurate arrangement of the objects (e.g., the way they would naturally fall into a box) is more important than perfect photorealism. 
3. Random Texture (The Boost to >95%): The decisive factor. By randomly varying surface textures, we force the AI to “understand” the pure geometry. This is the key to Physical AI. 

Quantity vs. Quality: The Efficiency Rule

Do companies need to generate millions of images? Our usualown experience from the experiments in the real-world lab suggests: No. 

Starting at about 10˝00 to 1,500 synthetic images per object class, the learning curve flattens out. More data adds little to no value. This means that high-performance AI is possible even without supercomputer infrastructure. 

The "Hybrid Boost" 

A tip from the Real-World Lab: Adding just 10 real photos ("10-shot") to a synthetic dataset can significantly improve accuracy. The combination of massive amounts of synthetic data and a few real-world references provides the optimal cost-benefit ratio.