Autonomous manipulator for depalletizing mixedly piled products of various kinds via object recognition including deep-learning-based boundary detection

This paper introduces a prototype autonomous manipulator for depalletizing mixedly piled products of various kinds in logistics centers that directly acts upon not only cardboard boxes but also wrapped objects such as plastic packages and bottle blocks. For this, we need object recognition with a high success rate on different types of objects and on tightly arranged objects of the same kind. Therefore, we considered the problem of determining boundaries between objects in rule-based image processing and designed an object recognition algorithm that combines deep-learning-based boundary detection and other elemental rule-based image processing functions. We also designed a robot gripper that can handle various objects safely and stably. In our test, where objects were closely arranged in pairs, the success rate of our algorithm was 24.9 percentage points higher than that of the purely rule-based technology. Using our algorithm, our prototype manipulator depalletized 48 products, including 24 different kinds, with a success rate of 97.5%. GRAPHICAL ABSTRACT