Using Actor-Critic Reinforcement Learning for Control and Flight Formation of Quadrotors

This paper introduces a near-optimal controller for the control of quadrotors. A quadrotor is described as a complex, twelve-state system. The paper simplifies the controller by considering it as two levels, the upper-level (kinematics) six-state controller and the lower-level (kinetics) twelve-state controller. An actor-critic optimal controller generates the desired velocities in the upper-level control, and its parameters are tuned by reinforcement learning. The desired velocities are generated using the upper-level controller, which is then used to solve for the lower-level control algebraically. Simulation results are provided to show the effectiveness of the solution.