To tackle the challenges in autonomous driving trajectory prediction, such as limited multimodal social interaction modeling, inefficient spatiotemporal dependency coupling, and poor adherence to vehicular kinematic laws, we propose SportsVAE—a dual-driven framework that integrates spatiotemporal modeling and kinematic constraints for vehicle motion trajectory generation. The model enhances interactivity and temporal correlation between features by incorporating additional neighbor features, which are processed using the Time Association with Neighbor Interaction Module (TANI). We introduce the Axis-aware Embedding Layer (AEL), an innovative replacement for the traditional embedding layer, which improves the modeling of entire trajectories and motion trend prediction through axial dynamic perception. Additionally, we propose Sports_loss, which incorporates a smoothness constraint into the loss function to effectively prevent the generation of non-physical trajectories. Extensive ablation experiments conducted on the NuScenes dataset validate the effectiveness of each method, achieving optimized results with a \({\text{minADE}}_{5}\) of 1.5562 and a \({\text{minFDE}}_{1}\) of 7.8762, improving endpoint prediction accuracy by approximately 9%.