Advances in Large Language Models for Robotics

Abstract

The application of large language models (LLMs) in robotics has garnered significant attention due to their potential to enhance robot understanding, reasoning, and interaction capabilities. This paper provides a comprehensive overview of the current state-of-the-art in integrating LLMs with robotic systems. We categorize and analyze existing methods from both task-oriented and model-oriented perspectives. Task-oriented applications include multi-object rearrangement, representation learning, error correction, and navigation, while model-oriented approaches encompass fine-tuning, direct use, and developing new specialized models. Furthermore, we propose a novel robot embodied intelligence framework that integrates LLMs with the Robot Operating System (ROS) to enable natural language understanding, task planning, execution, and user interaction. This framework aims to enhance robot autonomy and adaptability in unstructured environments, paving the way for new possibilities in real-world robot applications. We discuss the advantages and challenges associated with each approach and provide insights into future research directions in this rapidly evolving domain.