Waveform optimization for multi-input multi-output radar usually depends on the initial parameter estimates (i.e., some prior information on the target of interest and scenario). However, it is sensitive to estimate errors and uncertainty in the parameters. Robust waveform design attempts to systematically alleviate the sensitivity by explicitly incorporating a parameter uncertainty model into the optimization problem. In this paper, we consider the robust waveform optimization to improve the worst-case performance of parameter estimation over a convex uncertainty model, which is based on the Cramer---Rao bound. An iterative algorithm is proposed to optimize the waveform covariance matrix such that the worst-case performance can be improved. Each iteration step in the proposed algorithm is solved by resorting to convex relaxation that belongs to the semidefinite programming class. Numerical results show that the worst-case performance can be improved considerably by the proposed method compared to that of uncorrelated waveforms and the non-robust method.