Reducing the energy consumption and improving the robustness of a Wireless Sensor Network (WSN) are the main requirements for many industrial and research applications. The sensors usually use a routing protocol in order to deliver the sensing data to a Base Station (BS) which may be far away from the monitoring area. Many algorithms proposed in the literature compute the routing process by clustering the network and by designing new election mechanisms in which the cluster-heads are chosen taking account of the remaining energy, the communication cost and the density of nodes. However, they do not consider the connectivity to the BS, and assume that all the nodes or only few prefixed nodes are able to directly communicate with it. We believe that this assumption is not suitable for many applications of WSN and to tackle this problem we propose CESAR, a multi-hop and energy-efficient routing protocol for large-scale WSN which includes a new cluster-head selection mechanism aware of the battery level and the connectivity to the BS. Furthermore, our solution employs an innovative hybrid approach to combine both clustering and on-demand techniques in order to provide an adaptive behavior for different dynamic topologies. Simulation results show that our solution outperforms in terms of energy consumption and data delivery other known routing algorithms in the literature.