Multi-UAV Adaptive Cooperative Formation Trajectory Planning Based on an Improved MATD3 Algorithm of Deep Reinforcement Learning

Multi-unmanned aerial vehicle (multi-UAV) cooperative trajectory planning is an extremely challenging issue in UAV research field due to its NP-hard characteristic, collision avoiding constraints, close formation requirement, consensus convergence and high-dimensional action space etc. Especially, the difficulty of multi-UAV trajectory planning will boost comparatively when there are complex obstacles and narrow passages in unknown environments. Accordingly, a novel multi-UAV adaptive cooperative formation trajectory planning approach is proposed in this article based on an improved deep reinforcement learning algorithm in unknown obstacle environments, which innovatively introduces long short-Term memory (LSTM) recurrent neural network (RNN) into the environment perception end of multi-Agent twin delayed deep deterministic policy gradient (MATD3) network, and develops an improved potential field-based dense reward function to strengthen the policy learning efficiency and accelerates the convergence respectively. Moreover, a hierarchical deep reinforcement learning training mechanism, including adaptive formation layer, trajectory planning layer and action execution layer is implemented to explore an optimal trajectory planning policy. Additionally, an adaptive formation maintaining and transformation strategy is presented for UAV swarm to comply with the environment with narrow passages. Simulation results show that the proposed approach is better in policy learning efficiency, optimality of trajectory planning policy and adaptability to narrow passages than that using multi-Agent deep deterministic policy gradient (MADDPG) and MATD3.