Sports practice is essential for maintaining physical health, as well as for cognitive function, memory, and concentration. It also helps decrease anxiety and stress levels while developing teamwork and leadership abilities. With advances in science and technology, the application of artificial intelligence in sports is becoming popular among people and has many benefits. In this work, we introduce a comprehensive approach for action recognition in Vovinam, specifically targeting ten techniques in the “Nhap Mon Quyen” lesson. We created a new dataset dedicated to this work, consisting of 170 video clips recorded from 14 volunteers who demonstrated various levels of competency, ranging from beginners to intermediate practitioners and professional athletes, to maintain objectivity and diversity. We introduce a Bidirectional GRU with Temporal Attention mechanism evaluated via thorough 14-fold Leave-One-Subject-Out (LOSO) cross-validation. Across baseline comparisons, ST-GCN provides the best accuracy (0.951), followed by our Bi-GRU-Attention (0.925) and PoseConv3D (0.909). Bi-GRU-Attention achieves a distinctive benefit: interpretable temporal attention that shows which movement phases dictate classification decisions. This interpretability allows instructors to provide focused feedback for technique correction that is missing from black-box GCN techniques. This study presents the first benchmark for Vietnamese martial arts, demonstrating the accuracy-interpretability trade-off essential for practical applications in training.