From False Certainty to True Confidence: Pose Estimation Beyond Deep Learning

Robots estimate pose from sensors like lidar, radar, or sonar. Estimating the pose fast with high accuracy is essential for safe navigation. Errors may lead to crashes, destroying equipment, being stuck or lost in inaccessible areas. To handle unexpected situations, when no humans are around, robots must be able to assess their uncertainty, like a person that realizes they are lost. Unfortunately, the established pose estimation methods are either inaccurate or overconfident. Stein mixture inference is a new Bayesian inference method allowing accurate estimation with sound uncertainty assessment, but it has only been used for offline analysis. Robots need fast online analysis.  In this project in robotics and statistics, we extend Stein mixture inference to online problems and evaluate it for robot navigation. An uncertainty-aware pose estimation will enable robust control in unexpected situations, where the existing methods provide false certainty.