Risk-aware autonomous search and rescue with multiagent reinforcement learning.

Area of Science:

• Robotics
• Artificial Intelligence
• Reinforcement Learning

Background:

• Autonomous navigation in dynamic, high-consequence environments like search and rescue (SAR) missions is critical.
• Multiagent robotic systems face challenges adapting to changing conditions and adversarial risks.
• Deviations in learned actions and evolving environmental obstacles impact autonomous agent performance.

Purpose of the Study:

• To develop a risk-aware multiagent reinforcement learning approach for autonomous SAR.
• To mathematically formulate the autonomous SAR problem considering adversarial and environmental uncertainties.
• To evaluate the proposed approach in diverse hazard scenarios.

Main Methods:

• Formulated the autonomous SAR problem using a risk-aware multiagent reinforcement learning framework.
• Designed and implemented numerical experiments for evaluation.
• Employed a centralized training with decentralized testing (CTDE) paradigm.

Main Results:

• The risk-aware multiagent reinforcement learning approach demonstrated effectiveness in handling dynamic and uncertain SAR environments.
• The CTDE paradigm facilitated robust performance of autonomous agents under adversarial conditions.
• Experimental results validated the approach's adaptability to evolving obstacles.

Conclusions:

• The proposed risk-aware multiagent reinforcement learning approach offers a promising solution for enhancing autonomous navigation in SAR missions.
• Further research is needed to explore advanced adaptation strategies and real-world deployment.
• The study highlights the importance of risk-awareness in multiagent systems for high-consequence applications.