Prev-Talk | Next-Talk | All-Talks | Talks-Sorted | MOOS-DAWG'26

Talk-25: From Learned Coordination to Safe Deployment: Integrating Compliance into Learning-Based Multi-Robot Marine Autonomy

Everardo Gonzalez [1], Tyler Paine [2], Manuel Agraz Vallejo [1], Gaurav Dixit [1], Michael Benjamin [2], Kagan Tumer [1]

[1] Oregon State University, Autonomous Agents and Distributed Intelligence Lab [2] MIT Dept of Mechanical Engineering, Marine Autonomy Lab

Autonomous multi-robot teams are well-suited for marine missions that benefit from coordination, such as reef exploration, subsea infrastructure maintenance, or search-and-rescue operations. In such settings, team success is oftentimes measured according to sparse, long-term mission outcomes related to team performance (e.g., reefs discovered or swimmers rescued), while robots must simultaneously comply with immediate safety and regulatory norms such as COLREGs.

Cooperative coevolutionary algorithms can generate coordinated behaviors provided these sparse, long-term mission outcomes as feedback signals. However, integrating high-level mission objectives with real-time compliance remains a practical challenge: coevolved behaviors that optimize team performance do not inherently guarantee safe or compliant operation.

In this talk, we present a framework that leverages MOOS-IvP's behavior mixing capabilities to blend multi-robot behaviors learned through coevolution with compliance-focused behaviors that are prescribed based on regulatory constraints. The idea is that our robots can coevolve expressive behaviors in simulation to achieve a high-level mission objective, and we can later inject pre-programmed behaviors at runtime to ensure our robots incorporate compliance in the field. In our results, we show that our approach achieves high team performance while avoiding collisions on a collaborative swimmer rescue mission with up to 8 vehicles in a hardware deployment, and 12 vehicles in simulation. This work bridges learning-based coordination with operational norms, and is currently under review for conference publication.

Categories:

  • Collaborative Autonomy