Real-Time Observability-Aware Inertia Parameter Estimation for Quadrotors

This study focuses on the system identification of a quadrotor under an unknown payload with a time-optimal trajectory. A model-based control scheme should be utilized for a quadrotor, which does not necessarily guarantee robustness to model uncertainty. Thus, accurate system identification is required during flight. However, inertia parameter identification for the control scheme is vulnerable to sensor noise without a trajectory that produces rich data. We utilized Kalman filter (KF), which estimates the angular velocity, to reduce noise. In the process model of KF, the variance attributed to model uncertainty is derived, and the derived variance plays a pivotal role in adjusting Kalman gain. Recursive least squares (RLS) was utilized to identify the inertia parameter. However, all inertia parameters cannot always be observed with a time-optimal trajectory. Thus, this study proposes a criterion that distinguishes between observable and unobservable parameters and the correction law depending on the criteria. The correction law prevents RLS from correcting the unobservable parameters. We call this method the observability-aware RLS with KF. This study compared RLS, RLS with a low-pass filter (LPF), and observability-aware RLS with KF. While RLS with LPF shows a sharp increase or decrease in moment of inertia (MOI), center of mass (COM) offset, and sensor location, ours does not. RLS without any filtering has an inaccurate estimation of MOI and the height of the sensor location. However, our approach is sufficiently accurate to be applied to model-based control.