Efficient AI Agent Migration in Vehicular Metaverses: A Behavior Cloning–Regularized Diffusion Policy Optimization Approach

Vehicular metaverses are an emerging paradigm that integrates intelligent transportation systems with digital spaces, providing immersive services through AI agents that perform environment perception and decision-making. Due to the limited computing resources of vehicles cannot sustain the heavy inference load of large language models and visual language models, AI agents need to be deployed on high-performance RoadSide Units (RSUs). The high-speed movement of vehicles requires AI agents to dynamically migrate to different RSUs to ensure service continuity. To this end, we propose an efficient AI agent migration framework, and model the AI agent migration process as a partially observable markov decision process. Given the high dynamics of the migration scenario, we develop a Behavior Cloning-regularized Diffusion Policy Optimization (BCDPO) algorithm. The BCDPO algorithm uses a diffusion model to capture the multi-modal policy distribution and applies behavioral cloning terms during the policy optimization process, which enhances the efficiency of policy exploration and the stability of updates. Numerical results show that the BCDPO algorithm significantly improves the convergence speed of the policy compared with the baseline method and reduces the overall system latency of AI agent migration.