Analysis of a neural network model trained on synthetic data used to control UAVs

Abstract

In modern systems of unmanned aerial vehicles (UAVs), high requirements for the reliability and safety of performed maneuvers determine the need to use intelligent control algorithms. Trained neural networks are capable of providing adaptive response to external disturbances and dynamic changes in the environment. However, the limitations of real-world flights and the risks of accidents when experimenting with training data stimulate the development of synthetic modeling techniques. The purpose of the study is to analyze the effectiveness of a neural network model trained on synthetically generated data for UAV flight control. The article discusses an approach to training a neural network model for controlling an unmanned aerial vehicle (UAV) using camera images and clouds of lidar points. The training was performed on synthetic data collected in the AirSim simulator (Microsoft). A comparative analysis of three configurations was carried out: a model working only with images, a model with a camera and lidar trained on one scene (3,500 examples), and a model with a camera and lidar trained on 15,000 examples from four scenes. The developed model was tested on seven different scenes. The results obtained indicate a significant improvement in metrics with an increase in the size and diversity of the training sample. The results obtained confirm the validity of the synthetic learning approach for UAV control tasks, providing a high level of accuracy and speed. The main limitations are related to adaptation to unpredictable real-world factors and the need for further validation in field trials. The integration of reinforcement learning and the expansion of synthetic scenarios by simulating extreme weather conditions are promising.