Post-Disaster Multi-Robot Target Search Based on Improved Distributed Reinforcement Learning
Qilong Zhang, Yi Feng, Cheng Li · 2025
Post-disaster environments are highly dynamic and complex, with numerous unknown obstacles and flowing interferences such as strong winds and landslides. Moreover, the specific locations of rescue targets are typically unknown. This paper proposes a distributed reinforcement learning strategy network (F-IQN), based on target field strength information for multi-robot multi-target search in unknown environments without prior information on obstacle and target locations. This network utilizes target field strength signals from potential target areas to guide multi-robots through a two-phase target search process: first encouraging robots to reach the search area and then searching for target points based on field strength information. Additionally, a piecewise reward function is designed to divide the target search task into an area arrival phase and a target point search phase, significantly enhancing task completion effectiveness. Furthermore, this paper introduces a dynamic obstacle sensitivity strategy and designs an adaptive algorithm to automatically adjust the sensitivity level to obstacles in the environment, further improving search safety and success rates. Experimental results demonstrate that F-IQN with dynamic obstacle sensitivity outperforms traditional algorithms (such as DQN and DDQN) in search success rate, time efficiency, and energy consumption, especially maintaining superior performance when the search area expands.