Large language models support debugging and coordination in underwater robotics
A Layered View of LLM Support in Underwater Robotics
Robotics
Summary
Underwater robots face many challenges like limited communication and hard-to-see environments, making operations complicated. The authors suggest that large language models (LLMs) can help operators by interpreting robot data, diagnosing problems, coordinating teams, and assisting mission planning. They point out that these models are not yet reliable enough to control robots directly but can be a valuable support tool. Their idea is to use conventional robot controls for direct actions and LLMs for higher-level assistance during missions.
What this means in practice
- •For underwater robot operators: Assist operators by providing explanations of telemetry data and help diagnose faults during underwater missions.
- •For marine robotics coordination teams: Support coordination and replanning among heterogeneous underwater assets and teams under limited communication conditions.
A position paper. It proposes an approach and reports no results.
Authors
Federico Celi, Leonardo Zacchini, Gabriele Ferri
Abstract
This extended abstract discusses a specific and, in our experience, practically relevant role for large language models (LLMs) in underwater robotics. Our main premise is that, in underwater fielding missions, the immediate difficulty is often not the synthesis of new low-level control laws, but the operation of complex robotic systems under limited communications, partial observability, and expensive field time. In particular, a significant fraction of the effort during sea trials is devoted to fault diagnosis, telemetry interpretation, mission adjustment, and coordination across heterogeneous assets and operators. We argue that LLMs may be useful at this higher operational level, especially when grounded in structured vehicle state, alarms, logs, mission context, and existing planning or validation tools. At the same time, current limitations in grounding, consistency, and causal reasoning make these systems lack the predictability required to guarantee direct control, especially in real-world critical applications, although we do not exclude a future evolution in this direction. This viewpoint suggests a layered architecture in which conventional autonomy remains responsible for control, perception, estimation, and validated mission execution, whereas LLM-based modules serve as an additional support layer for debugging, operator interaction, constrained replanning support, orchestration, and post-mission analysis.