In this research, we develop a dialog system for a service robot which brings user-specified objects such as cans, bottles, and PET bottles from a refrigerator to a physically handicapped user. This paper describes two of the functions that are necessary for the system : one is the function to recognize objects in the image by using a dialog with a user; the other is the one to estimate a meaning of a word which is not registered. When the system cannot recognize specified objects automatically, it obtains necessary information by interacting with the user by speech, and tries to recognize the target object. When the system detects a word which is not registered, the system estimates its meaning by using a model of probability and registers it. We show the validity of these functions by experiments with real refrigerator scenes.
