In situations where toxic or dangerous airborne material is leaking, mobile robots equipped with gas sensors are a safe alternative to human reconnaissance. This work presents the Domain Knowledge Assisted Robotic Exploration and Source Localization (DARES) approach. It allows a multi-robot system to localize multiple sources or leaks autonomously and independently of a human operator. The probabilistic approach builds upon domain knowledge in the form of a physical model of gas dispersion and the a priori assumption that the dispersion process is driven by multiple but sparsely distributed sources. A formal criterion is used to guide the robots to informative measurement locations and enables inference of the source distribution based on gas concentration measurements. Small-scale indoor experiments under controlled conditions are presented to validate the approach. In all three experiments, three rovers successfully localized two ethanol sources.