<p>Video-based Action Quality Assessment (AQA) has become a key research area in fields such as physical education, rehabilitation training, and skill assessment. However, existing research has largely focused on Olympic sports such as diving and gymnastics, with insufficient attention paid to traditional martial arts movements, which possess deep cultural significance and complex movement characteristics. In contrast, Tai Chi are characterized by a slow, gentle rhythm, strong continuity, and sparse keyframes, making it difficult for current AQA methods to effectively capture their quality characteristics. Therefore, this paper presents TaiChi-18, a TaiChi video dataset, and designs a dual-branch modeling network (TCMS-Net) for Tai Chi action quality assessment. The dual-branch feature aggregator designed in this paper is better suited to the characteristics of Tai Chi actions. On the one hand, the Residual Multi-Layer Perceptron branch uses multi-layer residual blocks to progressively refine and enhance each time slice. It integrates the global temporal context of the entire video through global average pooling to form a video-level representation; on the other hand, the multi-scale adaptive weighted pooling branch utilizes multi-scale 1D convolutions to learn the weights of key segments across different time scales, enabling adaptive attention to key frame segments. Subsequently, the fused video-level features are mapped to the final action score via direct score regression. Finally, we conducted experiments on the TaiChi-18 and MTL-AQA datasets to validate the effectiveness of TCMS-Net. The results demonstrate that our model outperforms existing AQA methods on TaiChi-18; simultaneously, it achieves good generalization performance on the MTL-AQA dataset.</p>