Abstract
This paper is corned with the desired height control of the quadrotor under the framework of deep deterministic policy gradient with prioritized experience replay (PER-DDPG) algorithm. The reward functions are designed based on an out-of-bounds plenty mechanism. By introducing a generalized integral compensator to the actor-critic structure, the PER-DDPG-GIC algorithm is proposed. The quadrotor is controlled by a neural network trained by the proposed PER-DDPG-GIC algorithm, which maps the system state to control commands directly. The simulation results demonstrate that introduction of generalized integral compensator mechanism can effectively reduce the steady-state error and the reward has been greatly enhanced. Moreover, the generalization ability and robustness, with respect to quadrotor models with different weights and sizes, have also been verified in simulations.