Resumen
Due to the differences between simulations and the real world, the application of reinforcement learning (RL) in drone control encounters problems such as oscillations and instability. This study proposes a control strategy for quadrotor drones using a reference model (RM) based on deep RL. Unlike the conventional studies associated with optimal and adaptive control, this method uses a deep neural network to design a flight controller for quadrotor drones, which can map the drone?s states and target values to control commands directly. The method was developed based on a deep deterministic policy gradient (DDPG) algorithm combined with the deep neural network. The RM was further employed for the actor?critic structure to enhance the robustness and dynamic stability. The RM?DDPG-based flight-control strategy was confirmed to be practicable through a two-fold experiment. First, a quadrotor drone model was constructed based on an actual drone, and the offline policy was trained on it. The performance of the policy was evaluated via simulations while confirming the transition of system states and the output of the controller. The proposed strategy can eliminate oscillations and steady error and can achieve robust results for the target value and external interference.