In this work, we propose a methodical approach for creating an adequately annotated dataset for improved automatic detection of welding defects in TIG-welded Aluminum 5083 by using YOLOv8 model. This dataset was created from scratch using public domain images on Kaggle and includes six common defect types (Lack of penetration, Misalignment, Lack of fusion, Burn-through, and Contamination) as well as good welds with 200 handpicked images per class for balanced class distribution. The model was trained in Kaggle Notebook, which includes tuning YOLOv8 hyperparameters as 100 epochs, resolution 640, and batch size 16. Inference testing was validated by a hands-on manual review of the predicted bounding boxes to verify detection accuracy and classes. This model was also used to save the weights for deployment so that in real-time this application can be directly used for robotic welding systems. It discussed the combination of such detection models with robotic manipulators and communication systems to increase both efficiency and accuracy of industrial welding processes. Using robotic arms, the welding process can be monitored dynamically through interfacing trained YOLOv8 model thereby eliminating the defects, reducing wastage, and quality assurance. Moreover, advanced communication protocols promote efficient information exchange between the detection system and manipulators, fostering concerted decision-making processes. This study enables transparent methodology and practical outcomes for the development of automated defect detection for industrial robotics, connecting machine learning, communication systems, and robotic manipulators towards further research in industrial automation and quality control.