In this paper, we examine the use of autoencoders as an optimization tool for the construction of noncoherent space-time MIMO codes. In particular, we consider the quasi-static block fading channel, where the channel state information is not available at either the transmitter or the receiver, and changes independently between transmissions. Different from traditional constructions which aim to maximize an approximation of the minimum pairwise distance of the constellation, we use the autoencoder to directly target minimizing the probability of error. We show that this different optimization goal leads to constellations with more favorable pairwise distancesâ™ distribution and better error performance at low to medium signal to noise ratios where the minimum distance is not the limiting factor. Finally, we present simulation results showing that the constructed codes outperform traditional Grassmannian codes up to a signal-to-noise ratio of 20 dB using the traditional generalized likelihood ratio test detector.