Propeller thrust models have been widely discussed in fluid mechanics and many existing models can incorporate its dynamic changes with relatively high accuracy. However, most of current platforms adopt simplified version assuming static airflow condition due to the complicated model formulation. This simplified model contributes large modeling error especially in situations while vehicles exploring in cluttered environments, navigating outdoor with unstable wind flows, or maneuvering aggressively with high relative inflow speed.
To ameliorate the control performance of the vehicle, we formulate the problem into two separate parts for optimization, and then introduce an adaptive control framework with online propeller thrust model learning, which addresses the problem of unmodeled dynamics due to aerodynamic effects and modeling error on quadrotors. In order to adapt to instantaneous external airflow variation using the limited onboard computational resources, our online learning algorithm features fast computation for real-time model update, which is incorporated in the attitude controller onboard. This approach effectively improves the tracking performance and control efficiency in flight tests. Therefore, this methodology enables the possibility of eliminating unmodeled aerodynamic impacts in vehicle dynamics.
In the following figure, we can see the improvement of tracking performance in position:
and improved performance in attitude(yaw) using the approach proposed.
The mild but fast converging effects eliminate the unmodeled terms from external wind source, ground effect while landing, and the uncertainties in moment coefficient and Izz in our model.
This work is ready for submission as a technical report in CMU, RI.
Videos demonstrating flight tests for aerodynamic impact learning subjected to external disturbance(wind from fan) are shown in the following videos:
Shih-Yun (Laila) Lo 