Hybrid Deep Learning Framework for Real-Time UAV Navigation with Adaptive Smoothness Control

The objective of the study is to develop a reliable and energy-efficient urban UAV navigation system capable of overcoming
GPS degradation, localization errors, dynamic obstacles, and inefficient flight trajectories by leveraging cellular network
signals and advanced deep learning strategies. This research introduces a hybrid deep learning-based navigation system
to enhance UAV reliability through cellular network signals. A multi-objective cost function is designed to optimize the
uncertainty in motion, smoothness in trajectory, distance traveled, and the avoidance of collisions. A three-dimensional
geometry-based channel propagation model (3DGCPM) is used to reduce the amount of inter-cellular interference. The
federated meta multi-agent graph deep reinforcement learning (F2MGDRL) framework enables UAVs to adapt and learn in a
decentralized and distributed manner by combining federated learning to allow for privacy-preserving decentralized training,
meta-learning for fast adaptation to the changing environment in which the UAV is located. An adaptive enzyme action
optimizer (AEAO) module provides for efficient and smooth trajectories with minimal abrupt turns. Experimental results
show a success rate of 98.7% for completed missions and a latency of 45 s, and a significant improvement in the smooth and
controlled movement of the UAV’s trajectory. The system effectively navigates dynamic urban environments while minimizing
energy consumption and abrupt maneuvers. The integration of cellular signals with a deep learning approach and adaptive
control strategies mitigates the shortcomings of conventional GPS-based navigation systems and static navigation systems to
deliver improved reliability, stability during flight, and improved operational efficiency.
Keywords UAV navigation · Adaptive smoothness control · Cellular network integration · Enzyme action optimizer ·
Multi-agent deep reinforcement learning