Научная визуализация, 2026, том 18, номер 1, страницы 79 - 87, DOI: 10.26583/sv.18.1.07

Neural Network-Based Dynamic Grasping of Moving Objects with Robotic Manipulators

Авторы: Yin Cao^1,A, A.A. Boguslavsky^2,B

^A Lomonosov Moscow State University, Moscow, Russia

^B Keldysh Institute of Applied Mathematics of Russian Academy of Sciences, Moscow, Russia

¹ ORCID: 0009-0008-7577-2327, caoyin1995@gmail.com

² ORCID: 0000-0001-7560-339X, anbg74@mail.ru

 

Аннотация

We explore the use of the open‐source library Stable Baselines3 to implement reinforcement learning via deep neural networks for controlling a manipulator of grasping moving objects along a conveyor belt. Unlike static object grasping, this task requires accounting for dynamic factors, significantly complicating the process. We provide a detailed description of the physical-kinematic modeling of the manipulator in PyBullet .and the integration of both the manipulator’s and the moving objects’ parameters into the neural network for training. The results of this study demonstrate that the decision-making capabilities and autonomous behavior provided by reinforcement learning algorithms can be successfully applied to complex tasks, such as dynamic object grasping.

 

Ключевые слова: manipulator control, reinforcement learning, application of neural network models, physical simulation, PyBullet.