Scholarlink

In the collaborative search and rescue mission between underwater Autonomous Underwater Vehicles (AUVs) and divers, due to the dynamic nature of the marine environment, there may be errors in the historical prior knowledge of AUVs, which may lead to mission failure. Underwater human-computer interaction is an important method to address the problem of outdated historical prior knowledge. By enabling interactions between divers and AUVs, AUVs’ prior knowledge is enhanced, thereby improving decision-making performance. However, current underwater interaction is mainly achieved through limited types of gesture recognition or keyword matching in speech, which can transmit a small amount of information and have a low utilization rate of AUVs information. To address this problem, we propose a search and rescue algorithm (USAR-LLM4RL) that integrates real-time interaction information between divers and multiple AUVs through a Large Language Model (LLM) and Multi Agent Reinforcement Learning (MARL) framework. Firstly, the real-time speech from divers is converted into information segments, which are then passed into a probability generation engine to convert them into high target presence probability maps. In addition, the search and rescue mission is modeled as a Markov decision process. Finally, we propose the USAR-LLM4RL algorithm to enable AUVs to use and maintain high target presence probability maps during search. The simulation results show that the USAR-LLM4RL algorithm improves search efficiency and reduces energy consumption.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Multi-AUVs-Human Collaborative Search and Rescue Scheme Based on Large Language Models and Deep Reinforcement Learning

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Peijun Dong,
  • Hang Tao,
  • Yang Zhao

摘要

In the collaborative search and rescue mission between underwater Autonomous Underwater Vehicles (AUVs) and divers, due to the dynamic nature of the marine environment, there may be errors in the historical prior knowledge of AUVs, which may lead to mission failure. Underwater human-computer interaction is an important method to address the problem of outdated historical prior knowledge. By enabling interactions between divers and AUVs, AUVs’ prior knowledge is enhanced, thereby improving decision-making performance. However, current underwater interaction is mainly achieved through limited types of gesture recognition or keyword matching in speech, which can transmit a small amount of information and have a low utilization rate of AUVs information. To address this problem, we propose a search and rescue algorithm (USAR-LLM4RL) that integrates real-time interaction information between divers and multiple AUVs through a Large Language Model (LLM) and Multi Agent Reinforcement Learning (MARL) framework. Firstly, the real-time speech from divers is converted into information segments, which are then passed into a probability generation engine to convert them into high target presence probability maps. In addition, the search and rescue mission is modeled as a Markov decision process. Finally, we propose the USAR-LLM4RL algorithm to enable AUVs to use and maintain high target presence probability maps during search. The simulation results show that the USAR-LLM4RL algorithm improves search efficiency and reduces energy consumption.