Real-time Trajectory Optimization for Flexible wing-UAV Based on Reinforcement Learning

Flexible wing-UAV consists of a payload and a flexible canopy. It is widely used for the airdrop supply and aircraft recovery. However, combining the terrain constraint and control input management, it puts higher requirements on the real-time capability of the trajectory optimization. Thus, to improve the real-time capability, we present a novel trajectory optimization method based on deep deterministic policy gradient (DDPG). In this paper, a 3-degree of freedom (DOF) model is built by analyzing the data of the actual flight test. Then, the trajectory optimization methodology is designed based on the DDPG. Comparing with the traditional optimization method, the optimized trajectory of the proposed method does not need to be recalculated for the different initial values with the deep reinforcement learning method by building the optimization model previously. It shows large improvement on the real-time capability. Finally, the complete simulation results prove the feasibility and correctness of the proposed methodology.