1998 Volume 13 Issue 2 Pages 221-230
In this article, we discuss mobile robot navigation under linguistic instructions of a route to the goal in an unknown environment, aiming at man-robot communication by a natural language. It is assumed that the instructions given to a robot describe a route based on distinctive landmarks including intersections, for example, "turn right at the 2nd intersection, and ・・・". These linguistic instructions are composed from recognition of landmarks by the instructor using his/her own land-mark models, which give his/her own viewpoint to determine correspondence between each part of a route and landmark symbols available for instructions. In order to specify the route that is implied by the instructor in the environment correctly, the robot needs to have the landmark models identical to those of the instructor. But, as is shown in the previous works of pattern recognition, it is very difficult to define models that realize the same result of pattern recognition as humans, even if we assume only simple landmarks such as intersections in a two dimensional environment as those to be recognized. Instead of seeking such a complete model of each landmark, we propose to infer the route implied by the instructor from the result of landmark recognition based on usual landmark models that are easy to be given to the robot. The key idea is that it is possible to detect misunderstanding about each landmark in the instructions as contextual inconsistency between the instructions and the result of landmark recognition if the instructions are redundant to specify the route uniquely in the environment. We also propose to learn the evaluation of sensory data enough to recognize each landmark correctly from the result of navigation.
