We consider the problem of interference alignment (IA) for the K-user constant multiple-input multiple-output interference channel (K-user MIMO IFC) over correlated channels with imperfect channel state information (CSI). Recent performance evaluations show that most of the existing IA algorithms suffer serious sum rate degradations when the available CSI is imperfect. To deal with this issue, an uplink-downlink (UL-DL) Average-Mean-Square-Error(AMSE) duality is firstly established for the K-user MIMO IFC. Based on this duality, a robust IA algorithm is developed. Numerical results show that the proposed algorithm not only achieves better sum rate performance than other existing algorithms, but can also accommodate to the case when the perfect CSI is not available.