Adaptive beamformers are usually based on the premise that noise is stationary and Gaussian, and hence completely characterized by a covariance matrix. The standard approach then attempts to use measured data to estimate that covariance and form a set of beamformer weights that optimally rejects the estimated noise. This standard method may not be robust to mismatch between real noise and the noise estimate. In particular, low frequency passive sonar noise is usually nonstationary, owing primarily to the motion of the shipping noise sources. We investigate alternative robust noise models that allow for nonstationarity. One such model [W. A. Kuperman and F. Ingenito, J. Acoust. Soc. Am. 67, 1988–1996 (1980)] assumes a uniform probability distribution for surface noise sources. This produces an optimal beamformer that is robust in uniformly rejecting surface noise, no matter where sources occur on the ocean surface or how they move. The noise performance of this beamformer will be compared to standard adaptive approaches for array configurations proposed by ONR for the Acoustic Observatory Program. The high fidelity simulation model used for noise prediction includes nonstationary source motion and realistic range-dependent multipath propagation effects. [Work supported by DARPA.]