<p>In response to the problem in college-level table tennis education where the assessment of hitting quality relies on subjective judgment and lacks quantitative basis, this study proposes a spatiotemporal graph Transformer architecture based on skeleton guidance. It dynamically optimizes the joint connection relationship through adaptive graph learning, fuses local structure perception and global temporal modeling through the dual-stream encoding of spatial graph Transformer and time Transformer, and captures spatiotemporal coupled features through the cross-attention mechanism. Experimental verification shows that this method achieves significant performance improvement in the task of hitting quality assessment. Adaptive graph learning successfully captures the joint coordination patterns specific to the action, and the attention mechanism realizes the automatic focusing on key technical stages and decision moments. This study breaks through the fixed topological constraints and effectively integrates structural inductive bias and long-range dependency modeling, providing a quantifiable and traceable automated assessment tool for university physical education. It can realize real-time technical diagnosis feedback in a consumer-grade hardware environment and provides a feasible technical path for computer-assisted physical education.</p>