Програмний пакет для адаптивного навчання контролерів роборуки на основі нейромереж

Abstract

The article deals with the development of a software solution for the application of neuroevolution algorithms when creating a controller for controlling a robotic arm. The main principles of the neuroevolutionary approach for training neural network controllers in tasks requiring reinforcement learning are considered. In particular, the advantages of the adaptive approach are determined for a wide class of scenarios in which the working limb can work: implementation of stable grasping, positioning, manipulation of objects. It is noted that the final result of neuroevolution is an optimal network topology, which makes the model more resource-efficient and easier to analyze. The paper considers a software system that provides the developer with all the necessary tools for modeling the behavior of a robotic agent in environments of various levels of complexity: both two-dimensional and three-dimensional. In addition, the possibility of specifying the state of the agent not only as a set of data from sensors, but also as an image of the current environment from the camera is considered. According to the results of the experiments, the high efficiency of the search for the best solution using the NEAT algorithm is noted. It has been established that the proposed solution allows productively obtaining an effective policy in the form of a neural network, which has a minimal configuration, which will allow to increase the speed of the controller, which is critical for the operation of a real system. Thus, the use of a software solution for the adaptive development of a neuroevolutionary controller for solving tasks with a robotic limb allows to increase the efficiency of the learning process and obtain an optimal network topology.