In the context of multi-robot system and more generally for Technological System-of-Systems, this paper proposes a multi-UAV (Unmanned Aerial Vehicle) framework for SLAM-based cooperative exploration under limited communication bandwidth. The exploration strategy, based on RGB-D grid mapping and group leader decision making, uses a new utility function that takes into account each robot distance in the group from the unexplored set of targets, and allows to simultaneously explore the environment and to get a detailed grid map of specific areas in an optimized manner. Compared to state-of-the-art approaches, the main novelty is to exchange only the frontier points of the computed local grid map to reduce the shared data volume, and consequently the memory consumption. Moreover, communications constraints are taken into account within a SLAM-based multi-robot collective exploration. In that way, the proposed strategy is also designed to cope with communications drop-out or failures. The multi-UAV system is implemented into ROS and GAZEBO simulators on multiple computers provided with network facilities. Results show that the proposed cooperative exploration strategy minimizes the global exploration time by 25% for 2 UAVs and by 30% for 3 UAVs, while outperforming state-of-the-art exploration strategies based on both random and closest frontiers, and minimizing the average travelled distance by each UAV by 55% for 2 UAVs and by 62% for 3 UAVs. Furthermore, the system performance is also evaluated in a realistic test-bed comprising an infrastructure less network, which is used to support limited communications. The results of the test-bed show that the proposed exploration strategy uses 10 times less data than a strategy that makes the robots exchanging their whole local maps.