Pruning-Based Multi-Agent DRL for Efficient Twins Migration in Vehicular Embodied AI Networks

In intelligent transportation systems, the convergence of large language models and embodied artificial intelligence (AI) has led to Vehicular Embodied AI Networks (VEANs). In VEANs, the integration of digital twins results in Vehicular Embodied Agent AI Twins (VEAATs), which enables AI assistants for in-vehicle application services. Constrained by computational latency and limited on-board computing resources, autonomous vehicles (AVs) must offload computationally intensive VEAATs to Roadside Units (RSUs) for remote execution support with substantial resources. Considering the limited coverage of RSUs and the mobility of AVs, the migration of VEAATs among RSUs is crucial for ensuring continuous and high-quality immersive services. Nevertheless, high-density traffic exacerbates the workload imbalance among RSUs, increasing the risk of overload. This paper models the interaction between AVs and RSUs as a Stackelberg game to optimize bandwidth resource allocation for efficient task migration. We employ a multi-agent deep reinforcement learning algorithm to approximate the Stackelberg Equilibrium (SE) and address the workload imbalance issue. Specifically, we propose Path eXclusion-based Multi-Agent Proximal Policy Optimization (PX-MAPPO), which integrates the PX-based neural network pruning algorithm with the MAPPO algorithm. Numerical results demonstrate that this algorithm reduces the number of network parameters, decreases model complexity, and minimizes performance loss.