A cost-efficient vision-based robotic manipulation system for smart IIoT production lines

Abstract

A low-cost, vision-based robotic manipulation framework designed for intelligent IIoT-enabled production lines is presented in this research. To achieve reliable performance under various industrial settings, the proposed system presents an edge-aware perception pipeline that combines calibration-aware localization, monocular vision, and GAN-assisted augmentation for real-time robotic manipulation. The framework maintains excellent detection accuracy while drastically lowering hardware complexity and cost, in contrast to traditional methods that rely on cost-prohibitive depth-sensing hardware. To enable precise and dependable pick-and-place operations, a tightly connected perception calibration control architecture is created to guarantee accurate mapping from image-space observations to robot workspace coordinates. Real-time decision-making and low-latency inference are made possible by the system's deployment on an edge computing platform. A collection of about 900 annotated photos taken in various lighting scenarios, object orientations, and spatial arrangements is used for experimental evaluation. With a mean detection accuracy of 94.2% with low variance, the results show robust convergence behavior and better performance than baseline models. The system meets real-time industrial needs with an average inference time of about 35 ms per frame. Additionally, scalable deployment and smooth communication across dispersed production environments are made possible by integration with lightweight IIoT communication protocols. All things considered, the suggested framework offers a workable and scalable way to connect IIoT system integration, real-time robotic manipulation, and vision-based perception. The creation of a perception calibration control coupling architecture that synchronizes visual perception, spatial synchronization, edge inference, and robotic execution within a single IIoT production environment is the study's primary contribution rather than the individual adoption of current algorithms like YOLOv8 detection, GAN-based augmentation, or hand-eye calibration. Real-time closed-loop manipulation is made possible by this concept in industrial settings with limited resources.