Compared with the progress made on human activity classification, much less success has been achieved on human interaction understanding (HIU). Apart from the latter task is much more challenging, the main cause is that recent approaches learn human interactive relations via shallow graphical representations, which is inadequate to model complicated human interactions. In this paper, we propose a deep logic-aware graph network, which combines the representative ability of graph attention and the rigorousness of logical reasoning to facilitate human interaction understanding. Our network consists of three components, a backbone CNN to extract image features, a graph network to learn interactive relations among participants, and a logic-aware reasoning module. Our key observation is that the first-order logic for HIU can be embedded into higher-order energy functions, minimizing which delivers logic-aware predictions. An efficient mean-field inference algorithm is proposed, such that all modules of our network could be trained jointly in an end-to-end way. Experimental results show that our approach achieves leading performance on three existing benchmarks and a new challenging dataset crafted by ourselves. Code is available at: https://git.io/LAGNet.