AquaChat++: LLM-Assisted Multi-ROV Inspection for Aquaculture Net Pens with Integrated Battery Management and Thruster Fault Tolerance

2025-08-06

AI Summary

To address the limitations of single-ROV systems in aquaculture cage detection—namely, low energy efficiency, insufficient hardware fault tolerance, and poor adaptability to dynamic underwater environments—this paper proposes the first large language model (LLM)-enabled multi-ROV collaborative detection framework. The framework adopts a hierarchical architecture: an upper layer leverages LLMs (e.g., ChatGPT-4) for natural-language-driven mission planning and event-triggered dynamic replanning; a lower layer integrates trajectory tracking, real-time state feedback, battery-aware scheduling, and thruster-fault compensation. Crucially, this work pioneers the integration of LLMs into underwater multi-agent control, enabling semantic-level task understanding and fault-resilient autonomous decision-making. Simulation results demonstrate significant improvements: 28.6% higher inspection coverage, 21.3% reduction in energy consumption, and 92.4% mission completion rate under single-thruster failure—confirming strong robustness and practical engineering potential.

Technology Category

• Intelligent Robots: Multimodal Perception & Sensor Fusion
• Machine Learning: Large Multimodal Models (LMMs)
• Multiagent Systems: Multiagent Planning

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

Inspection of aquaculture net pens is essential for ensuring the structural integrity and sustainable operation of offshore fish farming systems. Traditional methods, typically based on manually operated or single-ROV systems, offer limited adaptability to real-time constraints such as energy consumption, hardware faults, and dynamic underwater conditions. This paper introduces AquaChat++, a novel multi-ROV inspection framework that uses Large Language Models (LLMs) to enable adaptive mission planning, coordinated task execution, and fault-tolerant control in complex aquaculture environments. The proposed system consists of a two-layered architecture. The high-level plan generation layer employs an LLM, such as ChatGPT-4, to translate natural language user commands into symbolic, multi-agent inspection plans. A task manager dynamically allocates and schedules actions among ROVs based on their real-time status and operational constraints, including thruster faults and battery levels. The low-level control layer ensures accurate trajectory tracking and integrates thruster fault detection and compensation mechanisms. By incorporating real-time feedback and event-triggered replanning, AquaChat++ enhances system robustness and operational efficiency. Simulated experiments in a physics-based aquaculture environment demonstrate improved inspection coverage, energy-efficient behavior, and resilience to actuator failures. These findings highlight the potential of LLM-driven frameworks to support scalable, intelligent, and autonomous underwater robotic operations within the aquaculture sector.

Problem

• Enables adaptive multi-ROV inspection in aquaculture net pens
• Addresses energy constraints and thruster faults in underwater operations
• Integrates LLMs for real-time mission planning and fault tolerance

Innovation

• LLM-assisted multi-ROV adaptive mission planning
• Two-layered architecture with real-time task management
• Thruster fault detection and compensation mechanisms