Research on Optimization Algorithms for Packaging VR Display Based on Multimodal Interaction Model and BERT

Virtual reality (VR) technology has significant application potential in information display, and multimodal interaction can further enhance users' immersion and interaction efficiency. The BERT model has powerful semantic understanding. However, current research on integrating the two and applying them to the optimization of VR display systems in the packaging industry remains scarce, and this direction also faces practical challenges, such as insufficient multimodal information fusion and low computational efficiency. The optimization framework combines BERT-based semantic encoding with aligned gesture and gaze features through adaptive multimodal fusion. The fused representation supports interaction-policy updates driven by user feedback and response latency. The experimental results show improvements in the measured performance of the optimized system integrating multimodal interaction and BERT: the immersion score has increased from 72.3 to 89.6, an increase of 17.3 points; the operation response delay has decreased from 320ms to 110ms, a reduction of 210ms; the scene switching fluency has increased from 68.5% to 92.1%, an increase of 23.6 percentage points. In terms of resource occupation and rendering efficiency, the average GPU load has decreased from 85.3% to 62.1%, a decrease of 23.2%; the peak memory occupation has decreased from 3.2GB to 2.1GB, a savings of 1.1GB; the scene rendering frame rate has increased from 45fps to 68fps, an increase of 23fps; the single interaction response time has decreased from 2.8 seconds to 1.2 seconds, a reduction of 1.6 seconds. In terms of model performance, the BERT model's AUC (Area Under the ROC Curve) is 0.911, higher than 0.896 for the CNN model and 0.903 for the BiLSTM model. In some experiments, after about 250 epochs of training, the accuracy is close to 90% and the loss is about 0.5. When the dataset size of the K5 sub-model is 6000, the AUC value is close to 1.0. This optimization method can effectively improve system performance and user experience.