Network Coordinate System (NCS) is an efficient and scalable mechanism to predict latency between any two network hosts based on historical measurements. Most NCS models, such as metric space embedding based, like Vivaldi, and matrix factorization based, like DMF and Phoenix, use squared error measure in training which suffers from the erroneous records, i.e. the records with large noise. To overcome this drawback, we introduce an elegant error measure, the Huber norm to network latency prediction. The Huber norm shows its robustness to the large data noise while remaining efficiency of optimization. Based on that, we upgrade the traditional NCS models into more robust versions, namely Robust Vivaldi model and Robust Matrix Factorization model. We conduct extensive experiments to compare the proposed models with traditional ones and the results show that our approaches significantly increase the accuracy of network latency prediction.