Prediction-Based Vision for Robot Control

Abstract

This article points out that the domain of robot sensing has much more structure than that of general sensing. A robot sensing system must operate within time and accuracy limits usually mandated by the application. It commonly does this by precomputing as much information as possible about the robot's environment and the objects in it, and storing this information as a model of the world. In most cases, this knowledge corresponds to statistical or structural methods of identifying objects in images, but is encoded in such a way as to be useful only for recognition or object location. As robot tasks become more complicated, this approach becomes less viable. A more general approach to modeling is required when task and path planning must be effected at runtime rather than fixed beforehand. And when unknown objects must be handled or when the environment becomes too complicated, the simple methods break down.