Scholarlink

This paper presents the design, implementation, and preliminary evaluation of a Human-Machine Teaming (HMT) framework for task planning of autonomous vehicles, developed in collaboration an industrial shipbuilding partner. The system integrates Large Language Models (LLMs) to support natural-language interaction, situational awareness, planning, and explainability during real-time missions. Rather than a rigid command pipeline, the framework enables mixed-initiative teaming: operators specify high-level intents and supervise execution while the agent surfaces constraints, proposes alternative plans, and explains decisions. The framework was validated in a simulated environment focused on missions for Unmanned Surface Vessels (USVs). Results indicate the system can interpret natural-language commands, adapt to under-specified instructions with 96% success in ambiguous scenarios, and detect infeasible requests before execution.

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

Integrating Large Language Models for Explainable Human-Machine Teaming in Autonomous Vehicles

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Jaime P. Pérez,
  • Alejandro Romero,
  • William Infante,
  • Benjamin Malafiej,
  • Francisco Bellas

摘要

This paper presents the design, implementation, and preliminary evaluation of a Human-Machine Teaming (HMT) framework for task planning of autonomous vehicles, developed in collaboration an industrial shipbuilding partner. The system integrates Large Language Models (LLMs) to support natural-language interaction, situational awareness, planning, and explainability during real-time missions. Rather than a rigid command pipeline, the framework enables mixed-initiative teaming: operators specify high-level intents and supervise execution while the agent surfaces constraints, proposes alternative plans, and explains decisions. The framework was validated in a simulated environment focused on missions for Unmanned Surface Vessels (USVs). Results indicate the system can interpret natural-language commands, adapt to under-specified instructions with 96% success in ambiguous scenarios, and detect infeasible requests before execution.