We examine methods to robustly determine the optical depth to reionization from future 21 cm datasets in a manner that is robust against detailed assumptions about the reionization process as encapsulated in simulations. In particular, we construct a training set of simulations using two different semi-numerical models, which use the same underlying density field realization and are matched so that they produce identical optical depth and ionization histories but differ in the detailed morphology of the 21 cm emission, and in their power spectra.  We focus the machine learning algorithm on extracting the volume ionization fraction of hydrogen from the simulations as the most robust observable quantity and the one for which 21 cm measurements will provide uniquely.  We demonstrate a neural network architecture that can robustly measure the ionization history in the presence of random noise and spatial filters like those which will affect the next generation of reionization measurements.