We consider the joint constellation design problem for the noncoherent multiple-input multiple-output (MIMO) multiple-access channel. By analyzing the noncoherent maximum-likelihood (ML) detection error, we propose novel design criteria so as to minimize the error probability. Our first criterion is the minimum expected pairwise log-likelihood ratio over the joint constellation. From an analysis of this metric at high signal-to-noise ratio, we obtain further simplified metrics. For any given set of constellation sizes, the proposed metrics can be optimized over the set of signal matrices. Using these criteria, we evaluate two simple constructions: partitioning a single-user constellation, which is effective for relatively small constellations, and precoding individual constellations of lower dimension. For a fixed joint constellation, the design metrics can be further optimized over the per-user transmit power, especially when the users transmit at different rates. Considering unitary space-time modulation, we investigate the option of building each individual constellation as a set of truncated unitary matrices scaled by the respective transmit power. Numerical results show that our proposed metrics are meaningful, and can be used as objectives to generate constellations through numerical optimization that perform better, for the same transmission rate and power constraint, than a common pilot-based scheme and the constellations optimized with existing metrics.