An Interactive Robot System Using Zero-shot Chain-of-thought Prompting

With advancements in generative AI, its integration with robotics is accelerating rapidly. In particular, the development of large language models (LLMs) presents new possibilities for facilitating effective human-robot interaction (HRI) in various fields. This paper proposes a natural language-based interaction system that integrates LLMs with robots, aiming to enhance human-robot communication. The system recognizes voice commands and employs the chain-of-thought (CoT) prompting to transform implicit or abbreviated human commands, which are natural language-based, into executable robot action (motion) sequences. Instead of using fine-tuning methods that require additional training of the model, this study adopts prompt engineering techniques to enhance system efficiency. Furthermore, our proposed system can adapt to various operational environments through zero-shot reasoning capabilities. Experimental results demonstrate that the proposed system achieves high task decomposition accuracy and interaction (executing the decomposed task) success rate of 88.8% and 85.8% respectively, validating the potential of LLMs in robot control systems.