Attention-Based Value Factorization Reinforcement Learning for Robot Swarms Cooperative Control

Autonomous functionality in multi-robot systems necessitates advanced coordination mechanisms for seamless task execution. In this paper, MAVFAC—an attention driven multi-agent value decomposition framework with actor-critic architecture—is introduced to improve collaborative decision making. To refine environmental feature capture, state representation is optimized through an upgraded encoding module. The decision engine further integrates an attention-guided value decomposition scheme within the actor-critic paradigm. End-to-end policy generation is realized by fusing both modules, yielding continuous and dynamically adaptive cooperative strategies. Validation experiments on static target navigation and cooperative pursuit tasks demonstrate MAVFAC's superior efficacy in multi-robot collaboration.