A Framework for Multi-Agent UAV Exploration and Target-Finding in GPS-Denied and Partially Observable Environments

Area of Science:

• Robotics
• Artificial Intelligence
• Remote Sensing

Background:

• Multi-agent remote sensing in GPS-denied and partially observable environments presents significant challenges.
• Efficiently locating survivors or surveying points of interest requires robust navigation and exploration strategies.

Purpose of the Study:

• To present a novel framework for multi-agent target-finding.
• To address the complexities of decentralized planning and control in unknown environments.

Main Methods:

• The framework separates planning and control problems.
• Planning utilizes decentralized multi-agent graph search with an online Partially Observable Markov Decision Process (POMDP) solver.
• Control employs Deep Reinforcement Learning (DRL) for local continuous-environment exploration.

Main Results:

• The proposed framework successfully enabled multiple agents to locate targets in large, simulated environments.
• The system demonstrated effectiveness despite unknown obstacles and environmental obstructions.
• The approach proved robust in complex, partially observable scenarios.

Conclusions:

• The combined POMDP planning and DRL control framework offers a viable solution for multi-agent target-finding.
• The approach is adaptable for various time-sensitive remote-sensing applications, including disaster response and hazardous environment surveys.