Among the battery state of charge (SOC) estimation methods, the Kalman-based filter algorithms are sensitive to the battery model while the neural network (NN)-based algorithms are decided by hyperparameters. In this paper, a hybrid approach composed of a gated recurrent unit (GRU) NN and an adaptive unscented Kalman filter (AUKF) method is proposed. A GRU NN is first used to acquire the nonlinear relationship between the battery SOC and battery measurement signals, and then an AUKF is utilized to filter out the output noise of the NN to further improve estimation accuracy. The hybrid method avoids the establishment of accurate battery models and the search for optimal hyperparameters. The data of dynamical street test and US06 test are used as training dataset and validation dataset, respectively, while the data collected from the tests under federal urban driving schedules and Beijing driving cycle conditions are taken as testing dataset. As compared with some hybrid methods proposed in other literature, the hybrid method has the best estimation accuracy and generalization for various driving cycles at different ambient temperatures. The root mean square error and the mean absolute error all are less than 1.5%, and the maximum absolute error is less than 2%. In addition, it also exhibits powerful robustness against the abnormal values of the battery signals and can converge to the true value in just 5 s.