Enhanced SAC-RNN-based collision-avoidance methods in UAV networks with leader–follower scenario

Abstract Unmanned aerial vehicles (UAVs) are a promising technology for 5G/6G. Compared with a single UAV, a UAV swarm can perform important tasks more efficiently due to its robustness, survivability, and redundancy. Therefore, the cooperation of UAV swarms has attracted great attention, and collision avoidance is one of the most important research topics because the UAVs may collide with each other during the mission. In addition, the leader–follower swarm structure for UAV operations can be used in mission planning to better avoid collisions. Soft actor-critic (SAC) is one of the off-policy deep reinforcement learning methods that improves the stochastic strategy according to the maximum entropy architecture and can handle a huge continuous state and action space, and long short-term memory (LSTM) and gated recurrent unit (GRU) models show remarkable performance in predicting the future state of the observed target. This work combines LSTM/GRU with the SAC model to avoid collisions in a UAV network with leader–follower scenario. Experimental results show that the proposed methods have the lowest collision rate ($50\%$), the smallest distance between the leader UAV and each follower UAV ($28\%$), and the shortest maximum distance between each pair of follower UAVs ($50\%$) compared with other existing methods.