Mobile opportunistic networks (MONs) are intermittently connected networks, such as pocket switched networks formed by human-carried mobile devices. Routing in MONs is very challenging as it must handle network partitioning, long delays, and dynamic topology. Flooding is a possible solution but with high costs. Most existing routing methods for MONs avoid the costly flooding by selecting one or multiple relays to deliver data during each encounter. How to pick the ''good'' relay from all encounters is a non-trivial task. To achieve efficient delivery of messages at low costs, in this paper, we propose a novel group-based routing protocol in which the relay node is selected based on multi-level cross-community social group information. We apply a simple group formation method to both historical encounters (social relationships in physical world) and/or social profiles of mobile users (social relationships in social world) and build multilevel cross-community social groups, which summarize the wide range of social relationships among all mobile participants. Our simulations over several real-life data sets demonstrate the efficiency and effectiveness of the proposed method by comparing it with several existing MON routing schemes.