End-to-End Deep Reinforcement Learning for Swarm UAV's in a Partially Observable Environment:Comparison with APF and Boids Algorithms with Dynamic Parameter Tuning

The autonomous navigation of swarms of UAVs in dynamic and complex environments poses a critical challenge in robotic research due to the limited adaptability of traditional Boids and Artificial Potential Fields (APF) algorithms with static parameters. In this study, an end-to-end Proximal Policy Optimization (PPO)-based deep reinforcement learning architecture is proposed for swarm UAVs, enabling the leader UAV to navigate independently using LIDAR data, while follower UAVs optimize their algorithm parameters in real time under POMDP constraints and Kalman filter noise. In comparative analyses conducted under identical neural network architectures, the APF algorithm exhibited "chattering"issues - causing chaotic thrust fluctuations and motor wear due to high prediction errors - while the Boids model was observed to stabilize thrust values at the 0.60 level by aligning with PPO's variance damping capacity. As a result, although both approaches converge to a similar reward band, it was determined that the Boids dynamics offer a much more sustainable infrastructure for autonomous swarm systems with real-world constraints in terms of actuator health and flight smoothness.