Author(s) :

Aiden Wang.

Volume/Issue :

Abstract :

This paper reviews recent deep learning methods aimed at reducing random errors in low-cost microelectromechanical systems (MEMS) inertial sensors for mobile robot dead reckoning. Accurate localization remains a critical challenge in GPS-denied environments, particularly for platforms that rely solely on accelerometer data. Four representative studies were selected based on architectural novelty and relevance to inertial-only dead reckoning. The review analyzes their denoising strategies, including Generative Adversarial Networks (GANs), Physics-informed Neural Nets, Wave-U-Nets, k Nearest Neighbors (kNN), as well as their performance across evaluation metrics such as Absolute Trajectory Error (ATE), Relative Trajectory Error(RTE), and Relative Rotation Error (RRE). Hardware platforms and tasks are also compared to assess generalizability. Findings indicate that research in range extension remains limited but suggest that generative architectures are promising for improving accelerometer signal reconstruction. Further, the lack of unified experimental datasets highlights an opportunity for standardization in future work. Overall, this review emphasizes that unified metrics and methods are key to advancing practical inertial-based dead reckoning in low-cost mobile robots.