Pixel-based behavior learning

摘要

In this paper we address the problem of learning behaviors for autonomous mobile robots. We particularly focus on methods which enable a human user to train a robot in its real destination environment without giving an a-priori model. Using complex visual input typical of real situations in office environments we show that very simple visual features can be used to represent the perception/action relation specific to a given behavior. From this point we propose a learning model relying on a statistical collection of two-pixels features for representing a behavior. We then present the experiments made on a real robot and propose extensions of the model for activeperception and behavior selection.