Delay-tolerant networks (DTNs) have the potential to interconnect devices in regions that current networking technology cannot reach. To realize the DTN vision, routes must be found over multiple unreliable, intermittently-connected hops. In this paper we present a practical routing protocol that uses only observed information about the network. We designed a metric that estimates the average waiting time for each potential next hop. This learned topology information is distributed using a link-state routing protocol, where the link-state packets are "flooded" using epidemic routing. The routing is recomputed each time connections are established, allowing messages to take advantage of unpredictable contacts. A message is forwarded if the topology suggests that the connected node is "closer" to the destination than the current node. We demonstrate through simulation that our protocol provides performance similar to that of schemes that have global knowledge of the network topology, yet without requiring that knowledge. Further, it requires significantly less resources than the alternative, epidemic routing, suggesting that our approach scales better with the number of messages in the network. This performance is achieved with minimal protocol overhead for networks of approximately 100 nodes.