Revisiting Uncertainty Propagation in Robotics

This paper revisits the problem of uncertainty propagation in robotic systems. Standard approaches use Kalman filtering, either Extended (linearized) (EKF) or Unscented Transform-based (UKF) Kalman Filters to handle state estimation of robotic systems. These dynamic filters operate on predict-correct cycle where the current state is predicted based on prior state, and then corrected based on current observations using calculated correction weight. These approaches provide both mean and uncertainty (covariance) of the state estimates. Alternative state estimation techniques include Monte-Carlo based particle filters. The system state estimate is often used in composition with other transforms to localize objects in a world frame. Often these internal transforms are treated as known nominal quantities, and their uncertainty is ignored. In this paper, we revisit the handling of errors and compare several representations to highlight limitations of current approaches.