TL;DR
This paper argues that Large Language Models (LLMs) should not be used for direct control of underwater robots, but rather as a "support layer" for human operators. By integrating LLMs into a layered architecture, they can assist with high-level tasks like fault diagnosis, telemetry interpretation, and mission orchestration, where their ability to summarize complex data can significantly reduce the operational burden of sea trials.
What Problem Does This Paper Address?
Operating autonomous underwater vehicles (AUVs) is notoriously difficult due to "field constraints": low-bandwidth acoustic communication, delayed measurements, and the high cost of ship time. During sea trials, a massive amount of effort is spent not on writing control laws, but on trying to understand why a robot is behaving a certain way.
Currently, operators must manually sift through asynchronous logs, alarms, and telemetry to diagnose faults or adjust missions. While there is a trend in AI research toward "end-to-end" autonomy—where a model handles everything from perception to motor commands—the authors argue this is too risky for expensive underwater assets. The gap they address is how to safely leverage the reasoning capabilities of LLMs to help humans manage complex robotic systems without compromising safety or predictability.
Methods
The authors propose a Layered Architecture for robotic intelligence. Instead of replacing the autonomy stack, LLMs are positioned as an advisory layer.
The core of the method involves:
- Grounding LLMs: Rather than letting LLMs generate free-form commands, they are grounded in structured data, such as vehicle state, alarms, machine-readable logs, and mission context.
- Bounded Assistance: The LLM acts as a supervisor or interface. It aggregates data into reviewable summaries and fault hypotheses but does not have the authority to execute actions.
- Tool-Based Interaction: The authors advocate for patterns where LLMs invoke explicit tools (like planners or validation modules) rather than generating raw code or direct control signals.
The specific roles for LLMs across different layers of the autonomy stack are detailed in the paper's central table:
Datasets
Not discussed in the paper. As an extended abstract/position paper, it focuses on architectural frameworks rather than training on specific datasets.
Key Insights and Ideas
- Diagnosis over Control: The most compelling use case for LLMs in this domain is fault diagnosis. LLMs can reconstruct chronologies from distributed, asynchronous events (e.g., a payload restart combined with a DVL bottom-lock loss) to help operators identify root causes faster.
- On-board vs. Top-side: The authors envisage a dual use. On-board, LLMs can act as "intelligent filters," identifying the most critical information to transmit over limited bandwidth. At the command center, they help operators scrutinize recovered logs.
- Explainability as a Bridge to Trust: Trust in maritime unmanned systems depends on transparency. LLMs can contribute to "autonomy explainability" by translating complex system states into natural language summaries.
- Constrained Replanning: Instead of open-ended search, LLMs should help operators choose between a small number of feasible, pre-validated alternatives determined by the robot's health and environmental conditions.
Key Results
The paper provides a qualitative roadmap and architectural blueprint rather than quantitative experimental results. The primary "result" is the definition of a safe, layered integration strategy that separates high-level reasoning (LLM) from low-level execution and safety interlocks (conventional autonomy).
Limitations and Future Work
The authors are candid about the current limitations of LLMs that prevent them from taking a more central role:
- Hallucinations and Grounding: LLMs often struggle with causal and temporal reasoning, which is critical when events are asynchronous and partially observed.
- Predictability: Current models lack the determinism required for safety-critical underwater operations.
Future Work: The authors suggest exploring emerging approaches for structured tool use and constrained inference (e.g., TypeSafe AI’s Jev) to make LLM outputs more reliable. They also emphasize the need to evaluate these systems using field-relevant metrics like "time-to-diagnosis" and "operator workload" rather than just standard NLP benchmarks.
Summary updated