Improving GNSS Positioning Using Deep Reinforcement Learning with Self-Supervised Learning Based Data Augmentation

The global navigation satellite system (GNSS) positioning can hardly meet the requirement of autonomous vehicle applications in dynamic urban environments due to the influence of complex processes and measurement noises. Recently, data-driven deep reinforcement learning (DRL) has been proven to model complex errors in dynamic urban environments to improve GNSS positioning accuracy. However, the remarkable performance of DRL is highly dependent on a large amount of training data. High-quality and available data is limited in urban environments due to problems like signal interruption and attenuation, leading to the model trained with such limited data being hard to generalize well across various environments. In this paper, we propose a novel DRL-based positioning-correction method with self-supervised learning-based data augmentation to improve the generalization in urban environments with limited data. Specifically, to address the problem of poor generalization caused by limited data, we employ a data augmentation strategy to generate sufficient augmented data for training. Considering using augmented data directly for training could encode redundant factors of variation from augmented data, leading to unstable training on the RL optimization process. To address this problem, we decouple the data augmentation from RL by strictly using original data for policy learning while using augmented data for auxiliary self-supervised representation learning to simplify the optimization. The auxiliary self-supervised representation learning aims to maximize the mutual information related to the task between original and augmented data while minimizing the redundant factors. Moreover, to achieve the accurate current states of the vehicle agent, an attention module is used to adaptively learn the importance of different input features, with a long- and short-term memory (LSTM) module used to extract historical information from the observations. We validate our proposed approach on real-world datasets with limited data, which demonstrates that the proposed method outperforms existing model-based methods and state-of-the-art learning-based methods in positioning accuracy with about a 13.0% and a 6.9% improvement respectively.