An end-to-end Flight Control Method for UAVs Based on MD-SAC

Deep reinforcement learning (DRL) allows unmanned aerial vehicles (UAVs) to learn control policies for tasks in complicated and unfamiliar environments, hence it is widely employed in the field of UAV flight control. However, the model and operational environment of UAVs are typically simplified, rendering them unrepresentative of the real world. Furthermore, using only a single sensory data to control UAV flight is difficult to realize autonomous decision-making of UAVs. In this paper, an end-to-end flight control method for UAVs based on multimodal data fusion and Soft Actor-Critic (SAC) algorithm is proposed, named MD-SAC. First, this paper constructs the UAV model that is basically consistent with the real physical model, and forms a UAV multidata fusion state space including UAV information, UAV and target information and UAV sensor sensing information. Then, the strategy of directly mapping the multimodal data fusion results to the UAV torque and thrust is proposed to construct an end-to-end UAV hierarchical control model, and the convergence of the control method is accelerated based on the empirical playback mechanism. The experimental results show that the UAV based on the MD-SAC algorithm can effectively complete autonomous trajectory planning and adapt to a variety of complex environments, and the performance is improved in terms of robustness and generalization compared with the PPO algorithm and the optimized SAC algorithm.