1: \begin{abstract}

2: Visual Inertial Odometry is a powerful tool in robot navigation. However when visual feature tracking fails, inertial integration becomes a critical weak point of VIO. Previous work has focused on learning a motion model from the IMU data which can be specific to the pattern of robot locomotion. In this work, we instead propose to train a network to predict the immediate evolution of the IMU biases. Our preliminary results show that learned bias predictions can improve state estimation in visually challenging situations. Our experiments demonstrate a 56\% reduction in drift rate in these situations. We also use this approach to enable in-motion initialization of VIO and show faster bias convergence leading to less error. 

3: \end{abstract}

4: