Adaptive beamformer methods have been used extensively for functional brain imaging using EEG/MEG surface recordings. However, the sensitivity of beamformers to model mismatches impedes their widespread application, in practice. In this study, we propose a state-of-the-art technique, termed robust minimum variance beamformer (RMVB), which enables adaptive beamformers to be robust against model non-idealities. RMVB yields robustness by quantifying different sources of uncertainty for all given nominal lead-field vectors corresponding to all voxels of the brain. These uncertainty regions, which are estimated as hyper-dimensional ellipsoids, are then directly incorporated in the equations to estimate the spatial-filter weights of beamformers to be subsequently used to perform the imaging. Although different sources of uncertainty may exist in practice, it is straightforward to estimate the uncertainty ellipsoids empirically by building several forward models for the source space. We conducted numerous computer simulations to evaluate the performance of our proposed RMVB technique compared to earlier versions of beamformers used for functional imaging. Our results reveal that RMVB can outperform conventional beamformers in terms of localization error, recovering source dynamics and estimation of the underlying source extents, when uncertainty in the lead field matrix is properly determined and modeled. Thus, RMVB can be used in various applications of source imaging such as determining the epileptogenic zone in medically intractable epilepsy patients or estimating the time-course of activity, which is a required step for computing functional connectivity of brain networks. Although RMVB requires extra computational power to estimate the uncertainty ellipsoids, these regions are estimated only once before performing the imaging. Furthermore, parallel computing can also be employed due to the scanning scheme incorporated in beamformers, to further accelerate the reconstruction procedure.