The use of multiple robots for exploration holds the promise of improved performance over single robot systems. To exploit effectively the advantage of having several robots, the robots must be co-ordinated which requires communication. Previous research relies on a fixed communication network topology, a single lead explorer, and flat communication. This thesis presents a novel architecture to keep a group of robots as a single connected and adaptable communication network to explore and map the environment. This architecture, BERODE (BEhavioural ROle DEcentralized), aims to be robust, efficient and scalable to large numbers of robots. The network is adaptable, the number of explorers variable, and communications hierarchical (local/global). The network is kept connected by an MST (Minimum Spanning Tree) control network, a subnetwork containing only the minimum necessary links to be a fully connected network. As the robots explore, the MST control network is updated either partially (local network) or globally to improve signal quality. The local network for a robot is formed by the robots that are within a certain retransmission distance in the MST control network. BERODE implements a hierarchic approach to distributing information to improve scalability with respect to the number of robots. The robots share information at two levels: frequently within their local network and less frequently to the entire robot network. The robots coordinate by assuming behaviours depending on their connections in the MST control network. The behavioural roles balance between the tasks of exploration and network maintenance where the Explorer role is the most focused on the exploration task. This improves efficiency by allowing varying number of robots to take the Explorer role depending on circumstances. The roles generate reactive plans that ensure the connectivity of the network. These plans are based on the imposition of heterogeneous virtual spring forces. Our simulations show that BERODE is more efficient, scalable and robust with respect to communications than the previous approaches that rely on fixed control networks. BERODE is more efficient because it required less time to build a complete map of the environment than the fixed control networks. BERODE is more scalable because it keeps the robots as a single connected network for more time than the fixed control networks. BERODE is more robust because it has a better success rate at finishing the exploration.