Extended Model-Based Learned Inertial Odometry

This research introduces a method for inertial odometry in quadrotors that enhances generalization to unseen trajectories by using a body-frame representation and incorporating torque dynamics into the learning process. The approach predicts short-horizon relative displacements, which are fused with IMU measurements in an Extended Kalman Filter. Experimental results show that this method outperforms previous state-of-the-art learned inertial odometry approaches for quadrotor pose estimation on unseen trajectories. Evaluations on four public quadrotor datasets, where whole trajectory shapes were held out during training, confirm lower velocity errors on every held-out set. The study also finds that torque input is most informative when derived from rotor signals. Code, data pipeline, and trained models are publicly available.