Energy Constrained Hierarchical Underwater Monitoring via Local Multi-Agent RAG

Underwater monitoring is severely hampered by limited battery life, high communication costs, and the difficulty of transmitting large volumes of raw data from remote locations. To address these challenges, the authors propose a hierarchical, multi-agent architecture designed for long-term autonomous operation. The system employs a master-satellite topology: ultra-low-power microcontrollers (MAX78000/MAX78002) act as 'sentinels' that continuously monitor visual and acoustic signals, while an NVIDIA Jetson Orin NX serves as the 'master' node, activated only when specific events occur or scheduled processing is required.

The architecture integrates a fully local multimodal pipeline. When the master node is triggered, it performs data ingestion, target extraction, and species identification using BioCLIP/OpenCLIP embeddings stored in a local ChromaDB vector database. A dedicated identification layer uses taxonomic centroids and supervised classifiers to categorize marine life. The system is managed by a LangChain-based multi-agent framework, which coordinates tasks such as query routing, energy management, and the generation of structured, researcher-ready reports. This design minimizes energy consumption by offloading simple detection to the sentinel layer and reserving high-performance compute for complex reasoning.

The system demonstrates that hierarchical processing can effectively bridge the gap between continuous, low-power sensing and high-fidelity scientific interpretation. By compressing raw data into structured knowledge and embeddings at the edge, the architecture significantly reduces communication overhead, making it feasible to transmit meaningful insights via bandwidth-constrained channels like satellite or acoustic links. The approach was validated through case studies in visual fish detection and marine mammal acoustic classification, showing that the system can maintain high accuracy while operating within strict energy budgets.