Autonomous vehicle motion control along a program trajectory based on machine learning
V. M. Grinyak, K. S. Shutov, Alexander S. Devyatisilnyi, Anton V. Artemyev · Neurocomputers · 2025
The article is devoted to the problem of controlling the movement of an unmanned device along a program trajectory set by a sequence of points. Such movement, which involves bypassing points, is characterized by a variety of possible trajectories. In classical control theory, the choice of a specific trajectory is associated with solving an optimization problem. At the same time, on the one hand, the formulation of criteria for such optimization is a problem, and, on the other hand, in problems of movement along a program trajectory, it is typically not necessary to find mathematically optimal solutions, but rather, solutions that are suitable for a particular practical case. The goal of the article is to develop a motion control system based on a program trajectory that operates according to the «similarity criterion», aiming to make the device's movement in automatic mode resemble that of a human-controlled device. The tool for implementing such a control system is a neural network trained on simulated or real motion data. The paper proposes a software and architectural solution for a motion control system based on the separation of command generation and command execution subsystems. The interaction between subsystems is managed through a message queue. The solution offers modularity and flexibility, allowing for easy adaptation and customization of individual components within the system. A model of the motion control problem has been developed, using the example of an unmanned vehicle-type device. A neural network configuration has been proposed for generating control commands, and a method has been suggested for training it using data received from a operator or generated as a result of simulation. Several examples of neural network training that lead to different device movements have been considered. These examples demonstrate that the system can adapt to various conditions and provide a level of control as close as possible to that of a human. The proposed approach to motion control could be used to develop control systems for various types of unmanned devices, ranging from ground vehicles to aerial and aquatic drones. These control systems can be applied in the automotive industry, logistics, agriculture, and other fields where different control modes are required.