Automating configuration is the key path to achieving zero-touch network management in ever-complicating mobile networks. Deep learning techniques show great potential to automatically learn and tackle high-dimensional networking problems. The vulnerability of deep learning to deviated input space, however, raises increasing deployment concerns under unpredictable variabilities and simulation-to-reality discrepancy in real-world networks. In this paper, we propose a novel RoNet framework to improve the robustness of neural-assisted configuration policies. We formulate the network configuration problem to maximize performance efficiency when serving diverse user applications. We design three integrated stages with novel normal training, learn-to-attack, and robust defense method for balancing the robustness and performance of policies. We evaluate RoNet via the NS-3 simulator extensively and the simulation results show that RoNet outperforms existing solutions in terms of robustness, adaptability, and scalability.