Task-Oriented Sensing and Covert Transmissions for Collaborative Multi-AUV Systems

TL;DR — A multi-agent reinforcement learning framework (SVR-MARL) for teams of autonomous underwater vehicles that learns distributed cooperative policies under realistic, covert acoustic-communication constraints. It matters because it moves MARL beyond idealized "perfect communication" assumptions toward real physical-layer limits in stealth-sensitive coordination.

• Problem: Covert underwater missions force AUVs to rely on passive observation (incomplete perception), while acoustic communication for information sharing suffers long delays, interference, low reliability, and exposure risk.
• Gap addressed: Prior communication-oriented MARL treats messaging as ideal info flow, and traditional comms optimization only targets link-level metrics — neither captures how perceptual information actually contributes to the cooperative task.
• Method: SVR-MARL ("Sensed Information Value Realization") uses practical information to quantify the task-utility of shared perception, learning distributed policies that balance task performance against communication/exposure costs.
• Evaluation: Demonstrated via a case study on covert multi-AUV cooperative localization and tracking, claiming improved task efficiency with reduced unnecessary communication and exposure risk (qualitative demonstration; no quantitative results provided in the abstract).