Collision-Avoiding Flocking With Multiple Fixed-Wing UAVs in Obstacle-Cluttered Environments: A Task-Specific...

Area of Science:

• Robotics
• Artificial Intelligence
• Aerospace Engineering

Background:

• Coordinated flight of multiple unmanned aerial vehicles (UAVs) is crucial for complex tasks.
• Developing decentralized, collision-avoiding flocking policies for fixed-wing UAVs in cluttered environments remains a significant challenge.

Purpose of the Study:

• To propose a novel task-specific curriculum-based multiagent deep reinforcement learning (TSCAL) approach.
• To enable decentralized flocking with obstacle avoidance for multiple fixed-wing UAVs.

Main Methods:

• Decomposing the flocking task into subtasks and progressively increasing complexity.
• Utilizing a hierarchical recurrent attention multiagent actor-critic (HRAMA) algorithm for online learning.
• Implementing model reload and buffer reuse for offline knowledge transfer between learning stages.

Main Results:

• TSCAL demonstrates superior policy optimality, sample efficiency, and learning stability compared to existing methods.
• Numerical simulations validate the effectiveness of the proposed approach.
• High-fidelity hardware-in-the-loop (HITL) simulations confirm TSCAL's adaptability.

Conclusions:

• The TSCAL approach offers an effective solution for decentralized flocking with obstacle avoidance in multi-UAV systems.
• The curriculum-based learning strategy significantly enhances the learning process and policy performance.
• The method is validated through both simulation and HITL testing, showing practical applicability.