Abstract:High-precision localization is the basis for mobile robots to perform upper-level tasks, and it is also the primary problem that directly affects the performance of other tasks. Localizability is a measure of the accuracy of localization. Estimation of localizability can enable the robot to avoid areas that are difficult to accurately locate itself and improve the success rate of other tasks. First, the map matching and dead-reckoning, which is generally adopted in localization has been analyzed. Based on that, different neural network modules are designed for the modeling of the map matching and the dead-reckoning respectively. As a result, a multi-module deep neural network model (MMN) consisting of a convolutional neural network (CNN) model, a long short-term memory neural network (LSTM)model and a multi-layer fully connected neural network (MLFC) model, is constructed. In this paper, localization entropy has been adopted to represent the localizability. So, entropy by the map matching has been predicted by the CNN model, entropy by the dead-reckoning has been predicted by the LSTM model and the fusion of the two entropy predictions is conducted by the MLFC model to achieve the prediction of the localizability. Simulation and experimental results has shown that the proposed method can offer a direct and accurate prediction of the localizability where the error of the prediction is less than 5%.