MMOCIC framework coordinates up to 12 marine robots with safety compliance

Collaborative marine robot missions—such as underwater exploration, infrastructure inspection, or search-and-rescue—face a fundamental challenge: they must balance team-level performance with strict safety and regulatory norms. Traditional learning methods struggle with sparse feedback signals and the need to adapt compliance on the fly. In a new paper on arXiv, researchers present Marine Multi-Objective Compliance-Integrated Coevolution (MMOCIC), a framework that decouples the learning of coordinated behaviors from the enforcement of prescribed norms. This separation allows coevolutionary algorithms to optimize team objectives (e.g., maximizing rescue rate) while a separate compliance layer handles collision avoidance and rule adherence, turning the mission into a tractable multi-objective optimization problem.

MMOCIC was validated on a collaborative swimmer rescue simulation with up to 12 autonomous vehicles. In a real-world hardware deployment, 8 vehicles successfully executed the mission while maintaining safety standards. The framework consistently outperformed baselines that either ignored norms or treated compliance as part of the learning objective. Key to its success is the insight that operational norms are prescribed, not discovered—so they should be integrated rather than learned. This work opens the door to safer, more scalable multi-robot autonomy in maritime environments, with immediate implications for defense, civilian search-and-rescue, and underwater infrastructure inspection.