Interval-based Robot Localization with Uncertainty Evaluation

Being able to provide trustworthy localization for a robot in a map is essential for various tasks with safety-related requirements. In contrast to classical probabilistic approaches that represent the uncertainty as a Gaussian distribution, we use interval error bounds for the uncertainty estimation of a localization problem. To tackle and identify the limitations of probabilistic localization uncertainty estimation, we carry out comparison experiments between an interval-based method and a factor graph-based probabilistic method. Different measurement error models are propagated by the two methods to derive the robot pose uncertainty estimates. Results show that the probabilistic approach can provide very good pose uncertainty when there is no non-Gaussian systematic sensor error. However, if the measurements have unmodeled systematic errors, the interval approach is able to robustly contain the true poses whereas the probabilistic approach gives completely wrong results.