Autonomous Navigation Command Generation System for Underwater Robots Using LLM

Abstract

Recent advances in artificial intelligence, particularly in natural language processing with large language models (LLMs), enable robots to perform complex tasks. Integration with ROS to simulate control room conversation enhances robot autonomy, allowing them to automatically navigate and determine optimal paths. This study demonstrates the feasibility of implementing a system where robots can generate their own navigation commands by integrating ROS and LLMs. To achieve this, a simulation environment was constructed, data processing and command generation processes were designed, and three methodologies were experimented with by applying ontology and prompt engineering. The experimental results showed that using a single GPT model with information filtering provided high efficiency and success rates in command generation, and higher versions of the GPT model delivered improved performance. These findings present a preliminary methodology that improves robot autonomy and efficiency in autonomous controls and establishes a foundation for further enhancing the performance of robot control systems.