Large Language Models in Robotics: Applications, Challenges, and Future Directions from the Perspective of Embodied Intelligence

Abstract

Large language models(LLMs) are beginning to influence embodied intelligence by improving robots’ ability to perceive their surroundings, follow instructions, and collaborate with people. This paper reviews recent progress at the intersection of LLMs and robotics, focusing on practical uses in manipulation, navigation, and human-robot interaction. At the same time, it highlights key difficulties, including grounding in the physical world, ensuring reliability and safety, and achieving efficient real-time operation. To address these gaps, the author discusses directions such as combining symbolic reasoning with neural methods and enabling robots to learn continuously through feedback. The study concludes that LLMs can strengthen embodied intelligence, but turning prototypes into robust real-world systems will require more efficient, adaptive, and human-aligned designs.