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

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Unmanned aerial vehicles (UAVs), specifically quadcopters, require precise control for position and trajectory tracking.
• Conventional control algorithms can be complex and may not always adapt optimally to dynamic flight conditions.

Purpose of the Study:

• To evaluate various neural network architectures as direct, standalone control algorithms for quadcopter position and trajectory tracking.
• To compare the accuracy and computational performance of different neural networks against specific control objectives.

Main Methods:

• Utilized an artificial potential field method to generate diverse trajectory datasets for training and validation.
• Evaluated single-layer regression networks, double-layer perceptron regression networks, deep neural networks (DNNs), and residual networks (ResNets).
• Assessed trajectory tracking accuracy using root mean squared errors and Pearson's correlation coefficient, alongside computational latency.

Main Results:

• Deep neural networks (DNNs) demonstrated superior trajectory tracking accuracy (RMSE 1.0830, R=0.9624) and stable flight in untrained scenarios.
• Simpler architectures like single-layer perceptrons offered significantly lower latency, suitable for real-time control despite minor accuracy trade-offs.
• ResNet architectures showed suboptimal performance in both accuracy and latency, underscoring the need for architecture selection based on application requirements.

Conclusions:

• Deep neural networks can serve as effective direct control algorithms for quadcopter position tracking, potentially replacing traditional methods when sufficient data is available.
• The study highlights a trade-off between accuracy and latency, suggesting tailored neural network selection for specific UAV control tasks.
• Neural network-based control offers a promising approach for achieving high precision, reliability, and computational efficiency in UAV applications.