In this paper, we develop a robust signal space separation (rSSS) algorithm for real-time magnetoencephalography (MEG) data processing. rSSS is based on the spatial signal space separation (SSS) method and it applies robust regression to automatically detect and remove bad MEG channels so that the results of SSS are not distorted. We extend the existing robust regression algorithm via three important new contributions: 1) a low-rank solver that efficiently performs matrix operations; 2) a subspace iteration scheme that selects bad MEG channels using low-order spherical harmonic functions; and 3) a parallel computing implementation that simultaneously runs multiple tasks to further speed up numerical computation. Our experimental results based on both simulation and measurement data demonstrate that rSSS offers superior accuracy over the traditional SSS algorithm, if the MEG data contain significant outliers. Taking advantage of the proposed fast algorithm, rSSS achieves more than 75 x runtime speedup compared to a direct solver of robust regression. Even though rSSS is currently implemented with MATLAB, it already provides sufficient throughput for real-time applications.