Commonsense knowledge extraction for tidy-up robotic service in domestic environments

Commonsense is one of the keys to enable human-robot communication in daily life scenarios. It is very difficult for a robot to do tasks ordered by a human without having some basic knowledge to understand the human's commands. This paper proposes a method to automatically build commonsense knowledge for the "Tidy-up" service, in which a robot is asked to take objects such as books, cups, dishes on a table to appropriate places automatically. We defined three object classes that are necessary for the service, namely "Washable"-objects that need to be washed, "Reusable"- objects that need to be stored for reuse, and "Trashable"-objects that need to be disposed of. For each object, multiple attributes were extracted from both the ConceptNet knowledge base and the Google search engine, and fed to classifiers to classify the object into the appropriate class. To evaluate the proposed method, output from classifiers were compared with the result from actual human. The result showed that the proposed approach is efficient in classifying objects and in providing object type as commonsense knowledge, hence, helping the robots to understand human intention and to provide intuitive service.