This paper studies robust beamforming design in a relay-assisted multiuser multi-antenna cognitive radio (CR) network that coexists with a primary radio (PR) system via opportunistic spectrum sharing (OSS). The robustness is with respect to imperfect channel knowledge at transmitter. Since the CR-to-PR interference needs to be carefully mitigated in OSS, the recently developed cognitive beamforming (CB) technique is employed by CR base station (BS), CR relay station (RS), and secondary receivers (SRs). Meanwhile, joint CR-BS and CR-RS transmit precoding is adopted to solve the weighted sum rate maximization (WSRM) problem. To cope with the non-convex problem, we apply the classic altering optimization process, namely the Blahut-Arimoto algorithm. Interestingly, it is shown that within each iterative step, we solely need to treat some semidefinite programming (SDP) subproblems with linear matrix inequalities (LMIs), which can be efficiently solved by standard convex optimization methods. Numerical results demonstrate significant performance gain of our approach over non-robust beamforming strategy.