Data efficient Deep Reinforcement Learning for robust inertial-based UAV localization

Abstract

Precise localization is a critical task for many UAV-based applications. Inertial Measurement Units (IMUs), which measure acceleration and angular velocity, are commonly used for UAV localization due to their low cost and small size. However, IMU-based localization is prone to accumulating errors over time, which can significantly impact the accuracy of the localization. To address this issue, we propose a data efficient Deep Reinforcement Learning (DRL) method that enables learning how to correct localization errors from IMUs. Our approach utilizes a novel data augmentation method, along with an appropriate “hint” loss that can provide additional supervision during the training process. As a result, the proposed method requires a very small number of real-world examples and can be implemented using widely available low cost RGB sensors, ensuring that it can be readily applied in a wide range of different applications. We demonstrate the effectiveness of the proposed method in both simulation and real-world UAV experiments. In comparison to traditional supervised and DRL approaches, the proposed approach allows for achieving more precise localization with fewer real-world examples, making it a practical tool for adapting DL-based localization models for UAV applications.