Pose Extraction from Sample Sets in Robot Self-Localization - A Comparison and a Novel Approach

For state estimation in robotics applications, especially for robot self-localization, the Monte-Carlo approach has been a popular choice in recent years. Since the original proposal of the approach, several modifications and improvements, e.g., for more intelligent resampling or for efficiently coping with the kidnapped robot problem, have been proposed. Nevertheless, the currently used approaches for computing a final result given the sample set bear several drawbacks. In this paper, we compare the most common techniques and propose a new approach that is computationally inexpensive and able to deal with multimodal distributions in self-localization scenarios. All results have been gained in experiments with a humanoid robot in a robot soccer scenario.