Hierarchical Map Generation Considering Place Concept and Object Recognition for Household Assisting Robot

In this paper, we proposed a method to generate hierarchical maps by reflecting the concepts of places, which are generated by using deep learning network Places205, and the recognized objects, which are recognized by using deep learning model YOLACT, onto a 2D laser map. The hierarchical maps are essential for aiming to efficiently accomplish tasks such as object grasping by household assisting robots. For example, in response to a task like "take me some drinks", using the generated hierarchical map, the robot moves to a place with a high possibility of finding drinks first, such as the concept of "kitchen", and then move to the object (e.g. the object of refrigerator) with a high possibility of finding the specific target (e.g. the water bottle). Simulation and real-world experiments are conducted to validate the effectiveness of the generated hierarchical map.