As the need for improving public safety and rapid intervention continues to grow, real-time violence detection in public has become increasingly important. The goal of this study is to develop a highly accurate and lightweight deep learning model for real-time violence detection, to integrate it into a social robot to promote peaceful behaviors among children, and enable educators to quickly intervene in aggressive situations in educational environments such as schools and kindergartens. To this end, we used two large and benchmark violence video datasets- RLVS and RWF-2000- and trained three architectures: Transformer, 3D-CNN, and CNN-LSTM. Our proposed CNN-LSTM model achieved the highest accuracy of 96.46% on the RLVS dataset with the lightest architecture, outperforming the other approaches. To test the generalization of the model in real-world situations, the model was implemented on the Taban social robot, which captures and analyzes real-time 5-s video to detect violent content. The detection process took 0.1 s. If the robot detects violent behavior, it expresses its concern facially and verbally to the users. Ten distinct scenarios, including both normal and aggressive actions, were designed, and each scenario was conducted through 5 participant groups in front of Taban in the Social and Cognitive Robotics Lab. This test achieved 90% accuracy on these 50 tests, which highlights the outstanding generalizability of our model to be implemented in real-time surveillance systems such as hospitals, schools, and kindergartens.