Deep and Decentralized Multi-Agent Coverage of a Target with Unknown Distribution

This paper proposes a new architecture for multi-agent systems to cover an unknown distributed target quickly and safely and in a decentralized manner. The inter-agent communication is organized by a directed graph with a fixed topology and We model agent coordination as a decentralized leader-follower problem with time-varying communication weights. Given this problem setting, we first present a method for converting the communication graph into a neural network, where an agent can be represented by a unique node of the communication graph but multiple neurons of the corresponding neural network. We then apply a mass-centric strategy to train time-varying communication weights of the neural network in a decentralized fashion. This implies that the observation zone of every follower agent is independently assigned by the follower based on positions of its in-neighbors. By training the neural network, we can ensure safe and decentralized multi-agent coverage control. Despite the target is unknown to the agent team, we provide a proof for convergence of the proposed multi-agent coverage method. The functionality of the proposed method is validated by a large-scale multi-copter team covering distributed targets on the ground.