MPG - Fast Forward Reasoning on 6 DOF Pose Uncertainty

Estimation of the pose of an object, or equivalently estimation of a rigid motion, is a prerequisite to many tasks in robotics, such as perception and manipulation. We want to estimate object poses, consisting of orientation and position of a target object, with 6 degrees of freedom (DOF). We use dual quaternions for the representation of poses and Mixtures of Projected Gaussians as probability density functions over all possible object poses. The framework can deal with widely spread density functions and provides closed form calculations of their fusion. Further, the framework allows for compositions of motion. In this paper we present two different improvements of our framework: an explicit treatment of uncertainties due to approximations and a faster integration over the 6-dimensional space of poses. Finally, the improved framework is demonstrated using different combinations of feature types. Topic: Research and Development / modeling, planning and control, Keywords: pose estimation, probabilistic inference, forward perception