Multi-input multi-output (MIMO) is an attractive solution to high data rate underwater acoustic communications (UAC) with limited bandwidth. However, the long delay spread characteristic of UAC channels makes the equalization challenging for those MIMO schemes. Although maximum likelihood sequence estimation (MLSE) is known to be optimum and is intensively used in many terrestrial wireless applications, it is not applicable for UAC channels since the computational complexity of MLSE increases exponentially with the number of channel taps. This paper presents an iterative reduced state MLSE scheme for MIMO UAC by taking advantage of the inherent line-ofsight feature of UAC channels. The proposed scheme spans the MLSE state space only from the dominant UAC channel taps and iteratively eliminates the residual inter-symbol interference (ISI) caused by the weak taps. Both Viterbi algorithm (VA) and sphere decoder (SD) are adopted in the proposed iterative MLSE. Analytical and numerical results are given to evaluate the performance of this scheme. Sea experiment results are also provided to further verify the performance of our proposed scheme. The proposed approach achieves 1.07% uncoded bit error rate (BER) with 2 transducers in SPACE08 sea experiment at a 1000m distance.