To address the complex challenges of multi-unmanned aerial vehicle (UAV) cooperative trajectory planning in unknown environments

including collision avoidance constraints and narrow passage traversing

an improved deep reinforcement learning method is proposed.This method introduces a long short-term memory (LSTM) network into the environment perception terminal of a multi-agent twin delayed deep deterministic policy gradient (MATD3) algorithm to enhance its ability to perceive dynamic environments.Simultaneously

a dense reward function based on the potential field is designed to improve learning efficiency and accelerate convergence.Furthermore

a hierarchical reinforcement learning training mechanism is constructed

comprising an adaptive formation layer

a trajectory planning layer

and an action execution layer

to systematically explore the optimal trajectory planning strategy.Simulation results demonstrate that the proposed method exhibits superior performance in terms of policy learning efficiency

trajectory planning optimality

and adaptability to narrow passages.The conclusion confirms that this method can effectively improve the robustness and adaptability of multi-UAV cooperative trajectory planning in complex and unknown environments.