Abstract: The integration of GNSS, inertial navigation system (INS) and visual sensors is an effective method to achieve high-precision navigation and positioning in complex environments. Aiming at the problems of reduced positioning accuracy and poor robustness caused by outliers of observed data in multi-sensor data fusion positioning in complex environments, this paper introduces an adaptive robust loss function in the state estimation of the graph optimization framework to adjust the residuals between the predicted values and the observed values, and performs weight reduction processing on larger residuals, effectively reducing the influence of outliers on state estimation. Improve the robustness and accuracy of positioning. Experiments were conducted in the indoor and outdoor switching environment and the urban vehicle environment. The results show that in the indoor and outdoor switching experimental scenario, the positioning accuracy of the method proposed in this paper is 19.5% higher than that of the random sampling consistency (RANSAC) method, 15.9% higher than that of the Cauchy loss function method, and 33.5% higher than that of the Huber loss function method. In the urban vehicle-mounted experiments, the positioning accuracy of the method proposed in this paper is improved by 92.6% compared with the random sampling consistency method, by 75.4% compared with the Huber loss function method, and by 5% compared with the Cauchy loss function method. In conclusion, in complex environments, the adaptive robust loss function has better outlier processing capabilities compared to traditional methods. It can effectively reduce the sensitivity of the positioning system to outliers, thereby improving the positioning accuracy of the system.