Autonomous Navigation and Control Method of UAV Swarm Based on Deep Reinforcement Learning

The efficient collaboration of drone swarms depends on real-time information sharing and communication between agents. However, in complex environments, communication delays, packet loss, and bandwidth limitations can affect the collaborative operations between agents and reduce the success rate of missions. This paper investigates the autonomous steering and command methods of drone groups based on profound retraining learning, focusing on the applicability of the multi-agent profound deterministic design gradient (MADDPG) engine in multi-agent restoration learn (MARL). First, this paper constructs an environmental model of the drone swarm, including the monetary function, movement space, and state dimension of every drone. The state space contains the drone’s position information, speed, acceleration, etc., as well as global information such as obstacles and target positions in the environment; the action space includes speed control and direction adjustment, and the reward function guides the behavior of the drone through factors such as approaching the target, avoiding obstacles and energy consumption. Then, based on the MADDPG algorithm, an algorithm framework suitable for this problem is designed, including the cognitive the agent’s value and policies. The policy network selects the best action based on the current state, and the value network is used to evaluate the value of the state, thereby adjusting the behavior of the intelligent agent. Through the strategy of centralized training and distributed execution, MADDPG can effectively solve the interaction and cooperation problems in multi-UAV collaborative navigation and achieve the optimization of autonomous navigation tasks. Experimental results show that in the case of three agents, the energy consumption of the MADDPG algorithm is 45J and the mission success rate is 94%.