Method for automatic object identification and robotic grasp point determination using machine vision

RELEVANCE of the study lies in the automation of manipulation operations involving complex-shaped objects in industrial production, which remains an important task in modern robotics. In many practical applications, objects are located on the working surface of a robotic system without overlapping but may have arbitrary orientations, which significantly complicates their automatic identification and the determination of grasp coordinates for a robotic manipulator. Although three-dimensional machine vision systems can solve this problem, they usually require RGB-D cameras and high-performance computing hardware, which significantly increases the cost and complexity of robotic systems. Therefore, the development of efficient methods capable of solving this task using only a monocular camera remains relevant. THE

Purpose

of the study is to develop and experimentally investigate a method for automatic identification of complex-shaped objects and determination of their grasp coordinates by a robotic manipulator using a machine vision system equipped with a single monocular camera.

Methods

Object recognition in digital images is performed using a convolutional neural network based on the YOLOv8 architecture, which allows detection of objects and estimation of their orientation on the working surface. The grasp coordinates are determined by analyzing the bounding box of the detected object and applying an experimentally determined offset of the grasp point relative to the geometric center of the object.

Results

Experimental evaluation was carried out using chess pieces of different shapes and colors as test objects. The object detection confidence was not lower than 0.9, while the processing time of a single image was less than 56 ms, which ensures real-time operation.

Conclusions

The proposed method enables automatic identification of complex-shaped objects and stable determination of their grasp coordinates using only a single monocular camera without expensive three-dimensional sensors.