This paper proposes a robust channel estimation strategy for two-way multi-antenna relay networks, where the two sources are not perfectly synchronized with each other. The relay is allowed to first detect the signal arriving order (SAO) and then estimate the channel matrix of source-to-relay link in order to construct the relay precoding matrix. In particular, the SAO detection is formulated as a composite hypothesis testing problem, and effectively tackled by using the generalized likelihood ratio testing (GLRT) method. Moreover, a two-step estimation algorithm for composite source-to-source channels is developed aiming at reducing the error probability of data detection, and the optimal training sequences to minimize the Cramér-Rao bound of the estimation mean square error (MSE) is derived. Simulation results show that the proposed channel estimation strategy can effectively mitigate the estimation error due to the asynchronous transmission, thus significantly outperforming the existing channel estimation method.