Smart cities can provide infotainment services such as virtual reality (VR) video for autonomous vehicles. However, stochastic variables in the system significantly degrade the Quality of Service (QoS) in current implementations. More effective handling of random factor interference becomes imperative for the ultra-low latency and QoS requirements of VR video and games. From the perspective of automobiles, the paper proposes a collective perception-driven stochastic programming framework for obtaining historical data, considering the effects of multiple stochastic variables in VR task offloading. For the first time, the Surrogate-Assisted Stochastic Programming (SASP) based on differential evolution algorithm with machine learning models is proposed to jointly optimize the task offloading time and cost while controlling the correlation of multiple stochastic variables. Through simulation and experiments, the results demonstrate the superiority of our proposed method. The SASP method can effectively utilize the historical data obtained by collective perception, which is conducive to achieving better results in uncertain environments and improving the reliability and stability of VR task offloading ( website ).