Early Pragmatic Language Development for an Infant Robot

摘要

Recent years have seen the growth of the idea in AI that machines can be designed as sociable creatures: that interacting with a robot could and should be as natural as talking to a fellow human and that machine learning could be achieved in a benevolent environment with a human teacher in ways that similar to children's learning. The two components of the future natural teaching of robots are a maximally natural interface and fast learning of categories and tasks from few examples. A natural interface lets humans intuitively use all the discourse cues and displays normally present in conversation. Sociable machines are embodied creatures which encourage natural interaction from humans and pick up on those cues and displays for the development of their own abilities, such as learning to perform new tasks, or learning to express themselves. This thesis describes a system that engages in such interactions and grounds it vocalizations in a manner inspired by the way human infants establish a communicative basis with their teachers and learn the meaning of words by engaging in functionally, pragmatically meaningful communication. It is a module of proto-linguistics behaviors designed and implemented on a robotic head that was already developed in the Artificial Intelligence Laboratory.