AAMAS Conference 2026 Conference Paper
Learning to Control Reconfigurable Multiagent Systems
- Manuel Agraz Vallejo
Reconfigurable multiagent systems (RMSs) are composed of individual agents capable of docking and undocking together to form composite agents with new capabilities that can adapt to diverse tasks. One such system, salp-inspired agents (simple thrust-based agents that can form chains for locomotion) have the potential for scientific monitoring in topologically intricate underwater habitats, such as caves, overhangs, and confined openings. Current approaches focus on controlling the locomotion of single salp units or small salp chains that are limited to operating on a fixed-size structure. However, this limits the scalability and robustness inherent to the salp chain design and requires new controllers for new chainconfigurations. Inourwork, weintroduceasetofgraph-based neuro-controllers whose structures directly map onto salp chains of arbitrary length. By representing salp units as nodes in a graph, we integrate the chain’s structure directly into the controller’s architecture, eliminating the need to redesign the controller whenever the chain grows or shrinks. Our results show that graph-based controllers retain up to 90% performance under zero-shot settings, demonstrating scalability and robustness to salp-unit failures.