Accurate position detection of objects with 2D cameras

Robots are getting smarter as Artificial Intelligence increasingly finds its way into industrial applications. But to work optimally, AI algorithms need large training datasets, and generating these datasets is very time-consuming and expensive.

To speed up this process, we have developed an end-to-end artificial intelligence workflow that recognises the object and estimates its position. Here, we start from synthetic data that we extract from an object's CAD file. With this, we realise a photorealistic training dataset and then train AI models for these tasks. The outcome of this AI model is a 2D vision system that lets robots perform various tasks, such as bin picking or assembly, without relying on expensive 3D vision systems. Our workflow allows engineers to set up vision-assisted robot applications faster. Moreover, this workflow is much more robust and faster compared to what would be possible when taking real photographs of an object. Instead of pulling thousands of photos and collecting them in training datasets, we now use synthetic data to generate a training dataset. This way, we can generate a dataset in a few hours, saving a lot of development time and cost.