Attitude estimation of a high-yaw-rate Mobile Inverted Pendulum; comparison of Extended Kalman Filtering, Complementary Filtering, and motion capture

This paper explores the accuracy of several state estimators used in Mobile Inverted Pendulum (MIP) robots. Accurate state estimation is essential for effective feedback stabilization of such vehicles, especially at high spin rates. The MIP estimation techniques compared in this work are the Complementary Filter, the Complementary Kalman Filter, the (proprietary) Digital Motion Processor (DMP) from the (common) TDK InvenSense MPU-9250, and a dynamically modelled Extended Kalman Filter (EKF). We also derive from scratch the equations governing the dynamics of MIPs undergoing high yaw rates, as used by the EKF, using a Lagrangian formulation. The MIP was then controlled at several different yaw rate setpoints, and the tilt angle estimates were compared with the (“ground truth”) measurements obtained via motion capture. Our test results indicate that the high yaw rate dynamic EKF and DMP are significantly more accurate than the usual Complementary Filter and planar dynamic EKF. The inaccuracy of the Complementary Filter is likely caused by the IMU not being aligned with the body's center of mass, creating a significant centrifugal force while spinning quickly.