Native cloud-based virtual reality (VR) games are shifting towards intelligent, real-time content generation. However, existing cloud action synthesis systems encounter challenges in terms of quality, efficiency, and adaptive evolution, which limits their capacity to meet stringent quality and latency requirements. In order to address this issue, a novel framework has been proposed. This is known as the Adversarial Procedural Action Synthesis Factory, the purpose of which is to generate high-quality, low-latency motion assets with runtime self-evolution capabilities. The proposed framework integrates a TCN-WGAN-GP model within a cloud-edge collaborative architecture, where encoding is offloaded to edge nodes and decoding remains in the cloud, thereby reducing end-to-end latency. The utilization of a digital evolutionary controller, founded upon the principles of CMA-ES, entails the integration of user feedback and environmental signals to ensure the perpetual optimization of generation strategies. Furthermore, a multi-granularity synthesis pipeline combining motion segmentation, style transfer, and anomaly repair enhances diversity and physical plausibility. The experimental results demonstrate a marked improvement in performance when compared to the baselines, with a mean opinion score of 4.35, action-text matching accuracy of 92.6%, and an average network jitter of a mere 3.75 milliseconds. The findings demonstrate the framework’s potential as an intelligent and scalable solution for self-evolving asset generation in cloud VR environments.
Xu et al. (Thu,) studied this question.