DIDO: Deep Inertial Quadrotor Dynamical Odometry

In this work, we propose an interoceptive-only state estimation system for a quadrotor with deep neural network processing, where the dynamics is considered as perceptive supplement of inertial kinematics. To improve precision multi-sensor fusion, train cascaded networks on real-world flight data to learn IMU kinematic properties, dynamic characteristics, and motion states along their uncertainty information, respectively. This encoded information empowers us address issues bias stability, dynamics, calibration during sensor fusion. The above multi-source fused into two-stage Extended Kalman Filter (EKF) framework better estimation. Experiments have demonstrated advantages our proposed work over several conventional learning-based methods.