Disturbance rejection and high dynamic quadrotor control based on reinforcement learning and supervised learning.

In this paper, we design and train a neural network controller for quadrotor attitude control to expand the application of quadrotors in more complex scenarios and challenging tasks. The neural network controller can allow the quadrotor to reject strong disturbance and realize high dynamic control. Because the quadrotor attitude control is a complex and high dimensional control problem, we propose a new framework that combines supervised learning and reinforcement learning (RL) to train the neural network controller. The neural network controller maps the states of the quadrotor to the control command of rotors in an end-to-end style. Besides, we propose the survival of the fittest principle for neural network preservation to obtain a better policy network during the RL training process. The numerical simulations demonstrate that: when the disturbance is more severe, the neural network controller trained by our method has better anti-disturbance ability than the proportion integration differentiation method and the incremental nonlinear dynamic inversion method, and the neural network controller supports high dynamic control to make the quadrotor achieves a large attitude angle. [ABSTRACT FROM AUTHOR]