Multi-UAV Cooperative Hunting in Obstructed Environments via a Multi-Agent Proximal Policy Optimization with...

Area of Science:

• Robotics and Autonomous Systems
• Artificial Intelligence
• Control Theory

Background:

• Cooperative hunting by unmanned aerial vehicle (UAV) swarms is crucial for autonomous decision-making in complex, obstacle-rich environments.
• Existing multi-agent deep reinforcement learning methods face challenges in exploration efficiency, environmental adaptation, and policy convergence for dynamic hunting tasks.

Purpose of the Study:

• To develop an effective algorithm for multi-UAV cooperative hunting of maneuvering ground targets in obstructed environments.
• To enhance the cooperative hunting capabilities of UAV swarms through a structured, progressive training approach.

Main Methods:

• Proposal of a curriculum learning (CL)-based Multi-Agent Proximal Policy Optimization (CL-MAPPO) algorithm.
• Implementation of a three-stage curriculum: basic obstacle avoidance/target search, environmental adaptability, and dynamic target hunting.
• Evaluation through ablation studies against standard MAPPO and comparative simulations against MADDPG and MADQN.

Main Results:

• CL-MAPPO demonstrated improved learning efficiency and achieved higher, more stable reward levels compared to vanilla MAPPO.
• The proposed algorithm exhibited a superior success rate in cooperative hunting tasks against baseline methods.
• Simulation results validated the effectiveness of the curriculum learning strategy in enhancing UAV swarm hunting performance.

Conclusions:

• The CL-MAPPO algorithm significantly improves multi-agent cooperative hunting performance in complex, dynamic environments.
• Curriculum learning provides a robust framework for training UAV swarms, overcoming limitations of traditional reinforcement learning approaches.
• The developed method offers a superior solution for autonomous multi-UAV coordination in challenging pursuit scenarios.