Steering Uncertainty: Mean-Field-Augmented Physics-Informed Reinforcement Learning for AUV Swarm Coordination in Underwater IoT
Year: 2026
Type: conference-paper
Abstract: Coordinating autonomous underwater vehicle (AUV) swarms for sensor data collection requires simultaneously maximizing spatial coverage and maintaining acoustic link quality, objectives existing multi-agent reinforcement learning methods treat as competing rather than complementary. This paper introduces MFA-PIWM-Diff, coupling a mean-field statistic for scalable O(N ) coordination, a physics-informed world model grounded in Thorp’s law providing channel predictions and an epistemic uncertainty signal, and a lightweight diffusion-based action refiner. The Kullback–Leibler divergence between Thorp priors and observation-updated posteriors enters the reward as a fixed-weight uncertainty penalty, alongside a soft augmented-Lagrangian penalty on the signal-to-noise ratio (SNR) floor; the floor is penalized rather than strictly guaranteed. Experiments with 32 AUVs over a 2 km × 2 km ocean domain (2,000 training episodes; deterministic evaluation over 90 episodes from three seeds) achieve 86.1% relay-based coverage at Rrel= 400 m (95% CI [83.4, 88.8]) and 75.4% strict coverage at Rcol= 300 m, with training-time mean SNR converging to 66.6 dB and the dual multiplier to λ ≈ 161.5. Across MADDPG, Greedy, PSO, and Random baselines, MFA-PIWM-Diff is the only method combining ≥ 85% relay coverage with a deterministic-evaluation per-step violation rate that, while still elevated in absolute terms (76.9%), is substantially below the unconstrained learned baselines (94–100%) and bounded by a maximum SNR deficit of only 5.9 dB versus 8.6–8.7 dB for failing ablations. The uncertainty penalty cuts the violation rate by 18 percentage points; the world model contributes 9.9 pp to strict coverage. Scalability from 4 to 32 AUVs maintains mean SNR above the 62 dB floor at every fleet size, and inference at 8.92 MFLOPs per agent decision executes within the 5 s control cycle on a laptop GPU.
Language: English
Subfield: Ocean Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204483333