Guilherme R.,New University of Lisbon |
Marques F.,New University of Lisbon |
Lourenco A.,New University of Lisbon |
Mendonca R.,New University of Lisbon |
And 2 more authors.
2016 IEEE International Conference on Systems, Man, and Cybernetics, SMC 2016 - Conference Proceedings | Year: 2016
This paper presents an incremental learning mechanism for context-aware switching between localisation methods which are available to the robots control system (e.g., GPS-based, map-based). The goal is to avoid the cumbersome and error prone manual mapping between localisation methods and environmental contexts. At each moment, the system determines which localisation method is performing best by comparison with the motion estimates produced by an odometer, assumed as accurate in the short-time. Then, the best performing method is associated to the current environmental context, which is defined by a novel descriptor built from the local occupancy grid. The result of this instance-based learning process is used online to estimate which localisation method performs the best in the current environmental context. The switching process is facilitated by the use of the de facto standard Robot Operating System (ROS) framework. The system was instantiated in a differential-wheeled robot equipped with a short-range 2-D laser scanner, and successfully validated on a set of field trials. © 2016 IEEE.