Integration of Artificial Intelligence (AI) with Factory Automation

Industry automation currently relies on manually programmed algorithms as well as precise measurements and controls to actuate factory robots that are resource intensive to set up. They are not easily adaptable to product changes, and they do not offer the required level of flexibility to support new features like predictive mantainance, run-time quality control, real-time production monitoring, and more advanced features. Additionally, shorter product development cycles require novel approaches that can now be provided by AI. We aim to design AI-enhanced controllers for manufacturing robots to support flexible production lines and product changes without the need for extensive reprogramming and replanning.

A promising research direction focuses on the integration of 6D pose detection into industrial production lines. 6D pose detection is the task of recognising an object’s position (i.e. translation) and orientation (i.e. rotation) in space. As part of this research, we are investigating how a 6D pose detection model can be safely and effectively integrated with the control software of an industrial robot. Such an integration would allow to guide the robot in real-time to precisely and safely pick up and move different objects within a manufacturing environment, without the need for reprogramming it.

We aim at creating a training database for 6D pose recognition models. This is needed since there are just a few examples of databases labeled for this kind of problem (most notably LINEMOD), but those databases have limitations (e.g. only one object per image) and, more importantly, there is no quick way to expand those databases to train an AI algorithm for different datasets. As experimental testbed, we will use a prototype device from AISmart which uses a FANUC robot and a camera to automatically generate labelled datasets. The database will be public and will be shared with the scientific community.

Industrial applicability: the proposed new database, which will include multiple objects and occlusions, will be used to test and enhance state of the art 6D pose models best suited for manufacturing precision and performance requirements. Finally, newly trained models will be tested in several “pick and place” scenarios typical of an industrial environment. Recently, 6D pose detection has seen significant progress in precision and performance reaching now 99% accuracy in real time (>20 FPS). Ultimately, we want to explore its industrial applications for production lines.