<![CDATA[This study aims to develop and evaluate the effectiveness of a transfer learning model on CNN with the proposed YOLOv12 architecture for recognizing hand gestures in real time on an integrated service robot. In addition, this study compares the performance of MobileNetV3, ResNet50, and EfficientNetB0, as well as a previously funded model (YOLOv8) and the proposed YOLOv12 development model. This research contributes to SDG 4 (Quality Education), SDG 9 (Industry, Innovation and Infrastructure), and SDG 11 (Sustainable Cities and Communities) by enhancing intelligent human–robot interaction for educational and service environments. The study applies an experimental method by comparing the performance of various transfer learning models in hand gesture recognition. The custom dataset consists of annotated hand gesture images, fine-tuned to improve model robustness under different lighting conditions, camera angles, and gesture variations. Evaluation metrics include mean Average Precision (mAP), inference latency, and computational efficiency, which determine the most suitable model for deployment in integrated service robots. The test results show that the YOLOv12 model achieved an mAP@0.5 of 99.5% with an average inference speed of 1–2 ms per image, while maintaining stable detection performance under varying conditions. Compared with other CNN-based architectures (MobileNetV3, ResNet50, and EfficientNetB0), which achieved accuracies between 97% and 99%, YOLOv12 demonstrated superior performance. Furthermore, it outperformed previous research using YOLOv8 (91.6% accuracy.]]>
