Until now, industrial gripping has been comparatively rigid: The geometry of the parts must be known, as must the exact pick-up and deposit position. A reliable handling process can be guaranteed on the basis of repeatable part feeds using fixed travel paths and the specification of target point coordinates. In the course of digitalization, the trend is now moving towards highly automated, fully networked and autonomous production systems.

In this context, the use of artificial intelligence (AI) is also becoming increasingly important. In conjunction with cameras, the first applications of cognitive intelligence in the gripper environment are already possible, enabling intuitive training by workers and independent completion of gripping tasks by the robot. Schunk is deliberately focusing on a practical, i.e. industry-oriented, design of the handling processes by limiting the number of component variations and thus streamlining the classification and training process.

In a first use case, which uses machine learning approaches for workpiece and gripping process classification, pluggable building blocks are combined as required and presented to a lightweight robot in any arrangement on a work surface for removal. In combination with 2D or 3D cameras, the self-learning system achieves a rapid increase in gripping reliability after just a few learning cycles: with every grip, the gripper learns how the workpiece can be successfully picked up and transported.

After just a few training rounds, the network classifies how to deal with the value pool of workpieces and the resulting combination options. In doing so, the gripper relies on learned empirical values on how to pick up and transport the workpiece. The intelligent performance of the algorithm is that future combinations and arrangements of the workpieces can be classified independently after just a short training period. This enables the system to handle parts independently and according to the situation. By continuously adapting the algorithms using AI methods, it is possible to identify previously unrecognized correlations and further refine the handling process. as