A knowledge-driven framework for Robotic Odor Source Localization using large language models

Abstract

Robotic Odor Source Localization (OSL) technology enables mobile robots to detect and navigate unknown odor sources in diverse environments. Traditional OSL methods, including bio-inspired, engineering-based, and machine learning-based approaches, face limitations of lack of adaptability to varying environments, significant computational resource requirements, and dependence on historical data. To overcome these challenges, we present a knowledge-driven framework that leverages Large Language Models (LLMs) to improve the robot’s navigation capabilities through contextual understanding and informed decision-making. A key feature of the proposed work is integrating an LLM agent with a memory module, which stores past experiences and recalls them during the decision-making process, allowing the robotic agent to make decisions based on current sensory inputs and previously acquired knowledge. Compared to traditional deep learning-based methods, such as Deep Q-Network (DQN), both simulation and real-world experiment results demonstrate that our framework significantly outperforms it in terms of accuracy, efficiency, and generalization across different environmental conditions.