Collision-Free UAV Flocking System with Leader-Guided Cucker-Smale Reward Based on Reinforcement Learning

Deep reinforcement learning has been applied to the control of flocking tasks for fixed-wing Unmanned Aerial Vehicles (UAVs) with successful results. However, previous research has given less attention to the design of flocking rewards, and the underlying mechanism of these rewards remains unclear. In this paper, we analyze the underlying mechanism of the flocking reward, and propose the leader-guided C-S reward to guide the fixed-wing UAV flock in a leader-follower structure, and prove that it is bounded when time is limited, which can avoid the gradient exploding problem. Additionally, we propose a collision-free fixed-wing UAV flocking system that uses multi-agent deep deterministic policy gradient to alleviate the non-stationary environment. The proposed system is simulate in 3 and 6 follower scenarios, and the results validate that it effectively controls UAV flocking.