METHOD OF ACCELERATED ADAPTATION OF UAV ON-BOARD NEURAL NETWORKS USING LARGE VISUAL-LINGUISTIC MODELS IN THE GROUND CONTROL STATION
Journal of Rocket-Space Technology · 20.04.2026 · 35 (1) · 79-86
DOI: 10.15421/452565 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Rocket-Space Technology, online erschienen am 20.04.2026. Verfasst von Illia Savchuk und 1 weiteren Autor:innen. Die Arbeit nennt 7 Literaturangaben.
Abstract (Originalsprache)
This paper solves the urgent scientific and applied problem of ensuring high operational deployment speed and adaptability of autonomous navigation systems for unmanned aerial vehicles in environments where global navigation satellite system signals are absent or jammed. For reliable computer vision-based autonomous navigation, semantic segmentation of the underlying terrain is critical. However, existing convolutional neural networks face the "domain gap" problem: when a UAV is moved to a new geographical area with specific terrain, recognition accuracy drops sharply. The classical approach to overcoming this problem involves collecting new satellite images and performing manual pixel-level annotation to retrain the model. This is a highly resource-intensive process that takes hours to days, making rapid UAV deployment impossible in dynamic conditions. To overcome this limitation, the study proposes a novel method for the accelerated adaptation of onboard neural networks based on the "Cloud Teacher / Edge Student" architectural concept. This concept is implemented within the software complex of a Ground Control Station. The core of the method involves utilizing powerful cloud-based Vision-Language Models capable of zero-shot recognition as the "Teacher". During pre-flight preparation, the operator uploads a new map of the target region to the GCS. The system automatically queries the VLM API to generate a ground truth semantic mask of the specified terrain, eliminating the manual annotation stage. The resulting data is instantly used for local fine-tuning of a lightweight and energy-efficient U-Net model with a MobileNetV2 encoder. Since fine-tuning occurs on a highly specialized micro-dataset with frozen basic feature extraction layers, the process takes only a few minutes on standard GCS hardware. In the final stage, only the updated weights of the lightweight model are uploaded onboard the UAV. This ensures that during the mission itself, the drone maintains full autonomy, requires no internet connection, and complies with the strict hardware constraints of onboard systems (SWaP-C). Analysis of the simulation results confirmed that the proposed pipeline reduces the preparation time of the navigation system for a new mission by a factor of tens, while maintaining high segmentation quality sufficient for the effective operation of computer vision algorithms and the elimination of navigation drift.
Zitieren
APA
Savchuk, I., & Halushka, Y. (2026). METHOD OF ACCELERATED ADAPTATION OF UAV ON-BOARD NEURAL NETWORKS USING LARGE VISUAL-LINGUISTIC MODELS IN THE GROUND CONTROL STATION. Journal of Rocket-Space Technology, 35(1), 79-86. https://doi.org/10.15421/452565
Vancouver
Savchuk I, Halushka Y. METHOD OF ACCELERATED ADAPTATION OF UAV ON-BOARD NEURAL NETWORKS USING LARGE VISUAL-LINGUISTIC MODELS IN THE GROUND CONTROL STATION. Journal of Rocket-Space Technology. 2026;35(1):79-86. doi:10.15421/452565
BibTeX
@article{Savchuk2026,
author = {Savchuk, Illia and Halushka, Yevhenii},
title = {METHOD OF ACCELERATED ADAPTATION OF UAV ON-BOARD NEURAL NETWORKS USING LARGE VISUAL-LINGUISTIC MODELS IN THE GROUND CONTROL STATION},
journal = {Journal of Rocket-Space Technology},
year = {2026},
volume = {35},
number = {1},
pages = {79--86},
doi = {10.15421/452565}
}
EUVIMED