DIVE: Deep Inertial-Only Velocity Aided Estimation for Quadrotors

This paper presents a novel deep-learning-based solution to the problem of quadrotor inertial navigation. Visual- inertial odometry (VIO) is often used for quadrotor pose es- timation, where an inertial measurement unit (IMU) provides a motion prior. When VIO fails, IMU dead reckoning is often used, which quickly leads to significant pose estimation drift. Learned inertial odometry leverages deep learning and model- based filtering to improve upon dead reckoning. Efforts for quadrotors, however, rely on sensors other than, or in addition to, an IMU, or have only been proven on a specific set of trajectories. The proposed generalizable approach regresses a 3D velocity estimate from only a history of IMU measurements, and the learned outputs are applied as a correction to an on- manifold Extended Kalman Filter. The proposed algorithm is shown to be superior to the state-of-the-art in learned inertial odometry. A 42% improvement in localization accuracy is shown over the state-of-the-art on an in-distribution testing set, and a 22% improvement is shown on an out-of-distribution testing set. Additionally, the proposed algorithm shows a 43% improvement over dead reckoning in VIO failure scenarios. Lastly, this paper is accompanied by an open-source implementation at github.com/decargroup/DIVE.