• Nikolay A. Mostakov, Trapeznikov Institute of Control Sciences of the Russian Academy of Sciences
• Alena A. Zakharova, Trapeznikov Institute of Control Sciences of the Russian Academy of Sciences

The article discusses the operation of robotic manipulation systems for the most popular tasks in the industry. The article provides an implementation of classical methods for grasping objects using a CAD model of the object, highlights their advantages and disadvantages. As a new solution, it is proposed to use a system based on the Action Chunking with Transformers (ACT) neural network architecture. The article details the use of ACT neural networks, the algorithm for training neural networks and launching them within the framework of technological operations of real production. The paper describes the hardware of the system, which includes the ARM95 Collaborative Manipulator, the RealSense Depth Camera D405 depth camera and the HTC VIVE position tracker. The following technological operations were considered as an experimental part of the work: grasping a box object, grasping a pencil object, painting a part and grinding a surface. The developed system shows that modern technologies, including machine learning methods, help to solve complex technological operations with a high level of productivity.

robot manipulator, cyber-physical systems, grasping objects, computer vision

2025-12-01