Abstract: In order to address the problem of autonomous coordination saturation attack of loitering munitions in urban scenarios, it is modeled as a decentralized partially observable Markov decision process (Dec-POMDPs). A specific reward function to ensure the arrival of loitering munitions at minimum time intervals and other reward functions with joint weight parameters are designed. Recurrent multi-agent deep deterministic policy gradient algorithm (R-MADDPG) is employed to train the policy for autonomous coordination saturation attack of loitering munitions. The success rate of several indicators is analyzed by Monte Carlo simulation method. The simulation and experiment results show that the mission success rate of autonomous cooperative saturation attack by the loitering munitions is 93.2% after training under the guidance of decision-making model, among which the mid-air collision avoidance rate between loitering munitions is 94.4%, the success rate of defense penetration in the air is 99.5%, the maximum time interval within 0.4 seconds is 95.3%.