One of the main challenges in cognitive radio (CR) communications lies in the system robustness to uncertainties. In this paper, we examine the impact of the noise power uncertainty on the performance of various detectors in CR networks. We consider both single and multiple CR nodes. For the single CR case, we compare the performance of the energy and likelihood ratio test (LRT) detectors in the presence of noise uncertainties. It is shown that both detectors perform about the same. We shall also investigate cooperative spectrum sensing based on soft-information combining, where the cooperating CR nodes experience different noise power uncertainties. We then propose a simple detection scheme that is more robust to noise variations and uncertainties than the conventional detection schemes. Numerical results are presented to verify the theory and demonstrate the robustness improvement of the proposed detection scheme.