Investigation of the USV-AUV Cooperative Environment via Reinforcement Learning and Its Impact on Data Collection and Energy Efficiency

Autonomous Underwater Vehicles (AUVs) and Unmanned Surface Vehicles (USVs) are key enablers for underwater sensing, yet their performance degrades under turbulent currents and wave-induced disturbances. We present a unified simulation framework for cooperative USV-AUV missions that operationalizes Fisher Information Matrix (FIM) within multi-agent reinforcement learning: FIM-guided USV positioning feeds into the agents’ observation and, when applicable, reward signals to reduce localization uncertainty and improve coordination. We evaluate three policies-Proximal Policy Optimization (PPO), Curriculum Reinforcement Learning (CRL), and Twin Delayed Deep Deterministic Policy Gradient (TD3) with FIM features (TD3-FIM)–under a single, reproducible setup across two regimes (normal and extreme sea states), with training and testing performed in both, using task-level metrics (data rate, total throughput), resource metrics (energy), system health (overflow events), and tracking error. A composite Reliability Index (RI) summarizes multi-objective performance. Results show that PPO consistently achieves higher reliability and more stable data collection than CRL in both regimes, while FIM-guided USV adaptation markedly lowers tracking error versus static baselines. The three-arm comparison establishes a practical benchmark for USV-AUV cooperation with physically motivated sea dynamics. Limitations include a simulation-based evaluation and simplified acoustic assumptions; future work will consider multi-USV coordination, latency/packet-loss models, and hardware-in-the-loop trials.