A Robot that Follows Natural-Language Directions
Prof. Andy Haas
February 12, 2008
If robots are to be useful in everyday life, they need to speak
English. This problem is not just a practical issue. It raises a
central question in cognitive science: how do we connect words to the
world around us? Most research on natural language processing is
words-in, words-out. The problem of connecting language to the world
does not arise. If robots are to follow directions in human language,
that is exactly the problem we must face.
I will describe a simulated robot that travels through a virtual
office building. Computer graphics allow humans being also to see this
virtual building. I have collected a corpus of sets of directions,
each tagged with the path it describes. This tagging provides strong,
detailed evidence about the meaning of the directions. On the basis of
this data, I have built a robot that follows directions.