Accurate estimation of the direction of arrival (DOA), direction of departure (DOD), and Doppler frequency is crucial in multiple-input multiple-output (MIMO) radar systems used for and/sea surveillance. Traditional nonadaptive estimation techniques, such as maximum likelihood and adaptive methods based on the minimum mean square error (MMSE), often underperform in land/sea clutter environments, which is essentially non-Gaussian and modeled as K distributed. Recent studies have explored the information-Theoretic learning (ITL) criterion as a robust alternative that leverages higher-order statistics to improve estimation under mild non-Gaussian conditions. However, ITL-based algorithms tend to have performance degradation in heavy-Tailed noises. To address this limitation, we propose a novel estimation framework based on the logarithmic hyperbolic cosine (LHC) criterion, a robust cost function designed to enhance resilience in adverse non-Gaussian noise scenarios. Building upon this criterion, we introduce the kernel LHC adaptive filter (KLHCAF), a non-linear estimation approach in reproducing kernel Hilbert space, that optimizes the LHC cost function for accurate recovery of DOA, DOD, and Doppler frequency. Simulation results in a practical MIMO radar context demonstrate superior performance of the KLHCAF filter over conventional ITL and MMSE-based adaptive methods, particularly under challenging non-Gaussian noise conditions.