Abstract
This research focuses on the development and evaluation of machine learning algorithms to enhance the navigation capabilities of unmanned aerial vehicles (UAVs). The main challenge addressed is ensuring reliable localization and autonomous trajectory planning in dynamic and GPS-denied environments. The study demonstrated that convolutional neural networks (CNNs) reduced localization errors by 18%, long short-term memory (LSTM) networks achieved 82% trajectory prediction accuracy with a 30% increase in stability, and Transformer models attained 89% test accuracy and 85% validation accuracy. Reinforcement learning (RL) methods further improved obstacle avoidance efficiency to 85% and achieved energy savings of 20%, although computational overhead increased by 30%. These outcomes are attributed to the integration of multimodal sensor data (LiDAR, IMU, GPS) and the application of deep learning architectures, validated through simulations in MATLAB/Simulink and Gazebo, as well as real-world testing using Raspberry Pi 4 and NVIDIA Jetson Nano platforms. A distinguishing feature of this research is the combined use of actual hardware prototypes and numerical models to verify the algorithms’ performance under real operating conditions. The results have practical relevance for military, environmental monitoring, and logistics UAV systems, especially in complex environments with variable lighting and dynamic obstacles.