Learning-Based Multi-UAV Flocking Control With Limited Visual Field and Instinctive Repulsion.

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Coordinated control of unmanned aerial vehicle (UAV) swarms is crucial for various applications.
• Traditional flocking algorithms struggle in dynamic and cluttered environments.

Purpose of the Study:

• To develop and evaluate a deep reinforcement learning (DRL) based flocking control policy for UAV swarms.
• To investigate the effectiveness of a centralized-learning-decentralized-execution (CTDE) paradigm for swarm control.
• To analyze the impact of limited sensory information and an encoded repulsion instinct on flocking behavior.

Main Methods:

• A CTDE paradigm was employed, utilizing a centralized critic network with augmented swarm information.
• Inter-UAV collision avoidance was implemented as an intrinsic repulsion function, not learned.
• The influence of varying visual field ranges on flocking control was simulated.
• Extensive simulations were conducted in diverse environments, including those with numerous UAVs, obstacles, and dynamic obstacles.

Main Results:

• The proposed DRL policy achieved high success rates: 93.8% in training, 85.6% with many UAVs, 91.2% with many obstacles, and 82.2% with dynamic obstacles.
• The inclusion of a repulsion function and limited visual field proved effective for flocking control.
• Learning-based methods demonstrated superior performance compared to traditional approaches in cluttered environments.

Conclusions:

• The developed DRL-based flocking control strategy is robust and effective for UAV swarms.
• Encoding collision avoidance as an instinct simplifies learning while maintaining safety.
• The approach shows significant promise for autonomous swarm operations in complex, real-world scenarios.