Aiming at the problem that the accuracy of the current compensation model for laser gyro bias error is low, an improved RBFNN bias error compensation model of laser gyro is proposed. The standardisation constant and data centre of the original data are obtained through the self-organising feature mapping network. The sample centre of the new sample data is obtained by the fastest decline of the expected variance of OLS algorithm. The results show, the improved RBF neural network algorithm has the best performance. under normal temperature, temperature change rate of 1C/min and temperature change rate of 3C/min, the zero-bias range of laser gyro is 3.491-3.508C/h, 3.992-4.021C/h and 4.092-4.123C/h, respectively. The research results provide new reference suggestions for the zero bias temperature compensation scheme of laser gyro at different temperatures.