Tenbrink, Thora & Reinhard Moratz

In a human-robot communication scenario involving groups of objects to be defined interactively, humans and robots need a shared reference strategy which suits the conceptual, perceptual and linguistic competences of both interaction partners. Where roughly similar objects are involved, one obvious method is to refer to the spatial position of the intended object. Here, qualitative descriptions like "to the left of" close the gap between the robot's ability and the human's inability to determine exact metrical positions. In our experiment, we tested human users' preferences for linguistic expressions of qualitative spatial reference when asked to instruct a robot to move towards one of several similar objects. To achieve this, we developed a computational model representing group based, salient-object based, and robot based projective relations. The robot used this model to interpret the spatial references.

The results indicated that those users who instructed the robot using a goal description rather than specifying minor actions, overwhelmingly referred to the group as a whole in order to identify the intended object by its position relative to the other objects in the group. Surprisingly, many users did not take into account that their instructions were ambiguous with regard to the perspective taken. We conclude, first, that group-based reference systems seem to be more salient for human spatial reference than has been acknowledged so far, and second, that dialogue modules suitable for comunication about space should be capable of asking contextually adequate questions to compensate ambiguities non-detected by the user.