Abstract: To address the issues of delayed feedback and difficulty in quantifying fundamental principles in traditional Tai Chi training, a Tai Chi motion quality evaluation framework, TaichiQualNet, is proposed. To resolve the computational redundancy and temporal inconsistency of conventional pose estimation algorithms when applied to slow, continuous Tai Chi movements, firstly a two-stage human pose estimation network is designed: Feature-Refined HRNet is used to efficiently extract 2D key points through motion saliency perception, while the Temporal Action Modeling network TaiFormer utilizes a Transformer structure to ensure the temporal consistency of 3D skeleton sequences. Based on this work, TaiBERT constructs an action recognition model based on an improved MotionBERT, which precisely analyzes core principles such as “Alternation of Emptiness and Fullness” through an attention mechanism. The model scientifically quantifies key metrics including center-of-gravity displacement, joint angle variations, and temporal continuity, and provides multi-dimensional scores and personalized feedback to trainees, thereby effectively compensating for the shortcomings of traditional instruction.