This work is based on the previous, “Propeller Model Online Learning”, and motivated to resolve the large modeling error particularly in situations while vehicles exploring in cluttered environments. Although the result of previous method seemed effective, it was not capable of adapting individual changes among different actuation systems on the quadrotor. In this work, since the model parameters that are not directly observable, they are predicted based on offline training using regression fitting a polynomial model on motor states. Due to the noisy environment of motor state estimation, online prediction along with state uncertainty propagation is introduced. This work is now in progress, and is going to be submitted to Robotics: Science, and Systems, 2015.
The fitted response surface can be shown in the following figure:
The polynomial model is introduced to adapt the parameters in the control input model. More robust performance in wind-disturbed environments is achieved through this methodology.