Painted vision and robot planning

Learning to cooperate in a bunpie microworld of 3/spl times/3 squares, each of two robots (robot A makes pies and consumes buns, while robot B makes buns and consumes pies) is rewarded if it sees the other robot "smile" (it "smiles" when it makes or consumes food). To help the robots predict and make plans in this world, their vision is augmented by a spatially structured, short-term memory of what they have seen during the past few steps. This "painted vision" is an economical, domain-specific way to improve prediction, and it is suggested that it is a step towards providing robots with that sense of "being in the world", which we enjoy.<<ETX>>