Reproduce “A Multi-Agent Reinforcement Learning Approach for Cooperative Air-Ground-Human Crowdsensing in Emergency Rescue” on OpenResearch

About this paper

Wenhao Lu, Zhengqiu Zhu, Yong Zhao, Yonglin Tian, Junjie Zeng, Jun Zhang, Zhong Liu, Fei-Yue Wang

Researchers developed HECTA4ER, a multi-agent reinforcement learning algorithm for coordinating heterogeneous entities (UAVs, UGVs, humans) in emergency rescue operations. The algorithm maximized the task completion rate, outperforming baseline methods by an average of 18.42% in simulations and demonstrating superior performance in a real-world case study.

Key insights

• A "Hard-Cooperative" policy for UGV-UAV interaction effectively manages UAV battery limitations, prioritizing replenishment to ensure continuous mission execution.
• The integration of global environment feature extraction (EIEM) and local agent decision-making based on historical observations (SEDM) is critical for robust performance in partially observable, dynamic environments.
• Modeling the task allocation problem as a Dec-POMDP and solving it with a CTDE multi-agent reinforcement learning algorithm enables effective coordination and maximized task completion under stringent time constraints.

Summary by alphaXiv.