Improved Simultaneous Localization and Mapping using a Dual Representation of the Environment

The designer of a mapping system for mobile robots has to choose how to model the environment of the robot. Popular models are feature maps and grid maps. Depending on the structure of the environment, each representation has certain advantages. In this paper, we present an approach that maintains feature maps as well as grid maps of the environment. This allows a robot to update its pose and map estimate based on the representation that models the surrounding of the robot in the best way. The model selection procedure is obtained by reinforcement learning and takes a decision based on the current observation. As we will illustrate in simulation as well as in real world experiments, this allows a robot to learn accurate maps in a more robust way than approaches using only feature or only grid maps.