Composable and Modular Autonomy for Maritime Robotics: Bridging Human-Robot Collaboration

Maritime operations incorporating multiple autonomous underwater vehicles (AUVs) pose serious challenges for human-robot cooperation due to environmental unpredictability and the complexity of handling multiple vehicles. This paper presents a novel modular and composable autonomy framework that improves human-AUV collaboration for marine missions such as multi-robot hull inspection. The structure employs a layered multi-agent system in which a supervisory agent, utilizing Large Language Models (LLMs), interprets operator commands, while separate LLM-driven robotic agents handle AUVs, negotiate task distributions, and oversee robot status. This advanced reasoning combines with Deep Reinforcement Learning (DRL) strategies for vehicle control (e.g., navigation, maintaining shared autonomy formation). Validation has been done via simulated ship hull inspection scenarios to demonstrate the framework's essential capabilities. The framework design, featuring collaborative target-tracking principles, seeks to lessen operator cognitive burden, enhance situational awareness, and facilitate adaptive mission operations in complex multi-AUV maritime systems.