Intelligence-Based Active Inference for Robots with Multimodal Large Language Models and Edge Computing

With the rapid development of artificial intelligence technologies, multimodal large language models (MLLMs) have demonstrated outstanding performance in handling various complex tasks. Robots, which are crucial for automated production, rely on path planning for efficiency and precision. Thus, applying MLLMs to the path planning of robots is a natural technological progression. However, the limited hardware in robots poses challenges for efficient and timely MLLMs-based inference. While traditional deep reinforcement learning (DRL) methods can offload these inference tasks to servers, existing DRL solutions often suffer from low data efficiency and are insensitive to users’ demands. To address these issues, this paper introduces an intelligence-based active inference (IAI) approach for offloading inference tasks and optimizing resource allocation in cloud-edge networks. Experimental results demonstrate that our proposed method outperforms mainstream DRL approaches in terms of data efficiency and better accommodates user dynamic requirements.