Improving Inertial-Based UAV Localization using Data-Efficient Deep Reinforcement Learning

Precise localization is a critical task for many Unmanned Aerial Vehicle (UAV)-based applications. Inertial-based navigation, which relies on Inertial Measurement Units (IMUs), is extensively used to this end, due to its low-cost and small footprint. However, IMU-based localization leads to accumulating significant localization errors. To overcome this limitation, in this paper we propose a data-efficient Deep Reinforcement Learning (DRL) method that enables learning how to correct localization errors from IMUs leading to more precise localization. In contrast with supervised approaches, the proposed method employs a novel data augmentation and regularization approach, which requires collecting a minimal number of real examples, while it is also platform-agnostic and can account for manufacturing impressions. The effectiveness of the proposed method is demonstrated both in a simulation environment, as well as using a real UAV.