MAMoE: A Multi-Agent Mixture-of-Experts Framework for LLM-Assisted 3D Object Reconstruction and Transmission

With the growing demand for high-quality 3D content in future wireless networks, minimizing energy consumption for 3D object reconstruction and transmission under strict performance constraints becomes a critical challenge. This paper investigates the problem of minimizing energy consumption for 3D object reconstruction and transmission in a network composed of multiple Network Service Provider-User Equipment (NSP-UE) pairs, where the Quality of Experience (QoE), the user information quality of service (QoS), and the power budget are considered as constraints. To solve the complex multi-pair hybrid task, a Multi-Agent Mixture-of-Experts (MAMoE) framework is proposed to simultaneously optimize 3D object reconstruction and transmission tasks. Integrating the strengths of Multi-Agent Systems (MAS) and Mixture-of-Experts (MoE), the proposed framework dynamically coordinates tasks among decentralized NSPs, with each provider leveraging specialized neural network experts for handling distinct reconstruction and communication subtasks. The reconstruction expert is responsible for selecting a subset of images to reconstruct the 3D object via classical Structure-from-Motion (SfM) and Multi-View Stereo (MVS) methodologies, where the reconstruction quality is evaluated through a comprehensive semantic assessment powered by a Large Language Model (LLM) module. The communication expert is responsible for transmitting the 3D object to the UE, which utilizes a tailored Proximal Policy Optimization (PPO) algorithm designed to optimize the transmission power. Simulation results demonstrate that the proposed framework outperforms both centralized MoE-Proximal Policy Optimization (MoE-PPO) and traditional decentralized Multi-Agent PPO (MAPPO) frameworks. Moreover, we investigate the impact of the number of NSP-UE pairs, QoE and QoS constraints, and time-varying channels on system performance.