Applying neural networks as direct controllers in position and trajectory tracking algorithms for holonomic UAVs

This study compares different neural networks as standalone control algorithms for position and trajectory tracking in holonomic UAVs, specifically quadcopters. The research's novelty lies in applying these algorithms directly for control. A position-tracking algorithm based on the artificial potential field method generated extensive training and validation datasets, simulating the tracked point's diverse trajectory shapes and velocities. The most popular neural network architectures were evaluated on the basis of their trajectory tracking accuracy and computational performance, i.e. single-layer regression networks and double-layer perceptron regression networks, deep neural networks, and residual networks. The results highlight that DNNs achieved the highest trajectory tracking accuracy, as measured by root mean squared errors (1.0830) and correlation coefficients (0.9624 given as Pearson's correlation) while providing satisfactory results and stable flight across untrained scenarios, in opposite to other neural networks. However, simpler architectures, such as single-layer perceptrons, exhibit significantly lower latency, making them suitable for real-time applications despite slightly reduced accuracy. In contrast, ResNet architectures underperformed in terms of accuracy and latency, emphasizing the importance of selecting architectures on the basis of specific control objectives. This study demonstrates that deep neural networks can directly control quadcopters, eliminating the need for conventional control algorithms for UAV position-tracking applications, provided sufficient learning data is available. The proposed approach ensures accurate trajectory tracking, effectively handling sudden turns while maintaining stable flight. These findings highlight the potential of neural networks for UAV control, balancing computational efficiency with high precision and reliability.