REINFORCEMENT LEARNING BASED LINEAR QUADRATIC REGULATOR FOR THE CONTROL OF A QUADCOPTER

In a practical implementation of the Linear Quadratic Regulator (LQR) for the control of a quadrotor drone, a key problem is the choice of the state and control weighting matrices. In this paper, we propose a Reinforcement Learning based LQR for quadrotor control. Leveraging the advances in deep reinforcement learning, Deep Deterministic Policy Gradient (DDPG) model is used to reset the elements of the Q matrix to achieve a faster response while minimizing the integral square error (ISE). Following the properties of the LQR control law, the LQR-DDPG controller is optimal and asymptotically stable. The proposed controller is compared with four other extensively used methods to choose the Q matrix. In the first method, that Q matrix is initially selected to be an identity matrix. Bryson's rule is used to set the Q matrix in the second method but not updated subsequently. Similar to the second method, the third method uses Bryson's rule to set the Q matrix but uses a proportional derivative controller along with LQR. The third method uses an iterative optimization algorithm that minimizes the integral square error (ISE) over the training trajectories to select the Q matrix. The simulation results show that LQR-DDPG is better than all benchmarking cases in terms of rise time, settling time and time of flight, all by a margin of 10% or more.