Teaching a Robot how to Spatially Arrange Objects: Representation and Recognition Issues

This paper introduces a technique to teach robots how to represent and qualitatively interpret perceived scenes in tabletop scenarios. To this aim, we envisage a 3-step human-robot interaction process, in which $(i)$ a human shows a scene to a robot, $(ii)$ the robot memorises a symbolic scene representation (in terms of objects and their spatial arrangement), and (iii) the human can revise such a representation, if necessary, by further interacting with the robot; here, we focus on steps i and ii. Scene classification occurs at a symbolic level, using ontology-based instance checking and subsumption algorithms. Experiments showcase the main properties of the approach, i.e., detecting whether a new scene belongs to a scene class already represented by the robot, or otherwise creating a new representation with a one shot learning approach, and correlating scenes from a qualitative standpoint to detect similarities and differences in order to build a scene hierarchy.