I. Introduction
In Recent years, visual-inertial odometry has seen tremendous progress (e.g. [1]–[4]), driven by the many potential applications of such technology for augmented/virtual reality and autonomous robots, in particular flying robots. Surprisingly, for wheeled robots, VIO is not trivial to employ due to observability limitations for planar linear motions [5]. This can be alleviated by integrating motion model constraints into the state estimate. A popular approach in the literature is to use wheel odometer measurements to this end (see e.g. [5]–[7]).