Moving from augmented to interactive mapping

Recently1 there has been a growing interest in human augmented mapping[1, 2]. That is: a mobile robot builds a low level spatial representation of the environment based on its sensor readings while a human provides labels for human concepts, such as rooms, which are then augmented or anchored to this representation or map [3]. Given such an augmented map the robot has the ability to communicate with the human about spatial concepts using the labels that the human understand. For instance, the robot could report it is in the ”kitchen”, instead of a set Cartesian coordinates which are probably meaningless to the human. Even if the underlying mapping method is perfect, two main problems occur in the approach of augmented mapping. When guiding a robot through a number of rooms, humans tend to not provide labels for every visited room [4]. The result is that the robot has difficulty to model where one room ends and the other room starts. This problem could be solved by detecting room transitions through the sensor data. Although good attempts using such an approach have been made in office environments [5, 6], applying these to other environments such as real homes is nontrivial. Another problem is that the generalization of the labeled map to newly acquired sensor data can be much different from the humans ideas. That is: there is a mismatch between the human representation and the robots representation. In our case the robots generalizes labels using visual similarities, while humans could use the function of the room. Even among humans there are differences between spatial representations. Think of a living room with an open kitchen. Where does the living room end and the kitchen begin? Our solution to both of these problems is to use pro-active human robot interaction. We briefly describe how the robot learns a map of the environment using a vision sensor and active dialog with a human guide. The method is implemented on Biron (the Bielefeld Robot Companion), see Figure 1, which supports an integrated human robot interaction system based on XCF (XML Communication Framework) complete with person attention, spoken dialog, person following, gesture recognition and localization components [7].