This paper presents a novel machine learning method to improve the accuracy of odor source localization (OSL), a core function for autonomous systems operating in complex environments. Existing OSL methods suffer from ambiguity problems, where robots incorrectly attribute odors to the wrong objects due to limitations in olfactory datasets and sensor resolution. To address this, we propose a novel machine learning method using diffusion-based molecule generation. This method expands the chemical space beyond the limitations of existing olfactory datasets and training methods, enabling the identification of previously undocumented potential odor molecules. The generated molecules can be more accurately validated using advanced olfactory sensors, detecting more compounds and informing better hardware design. By integrating visual analysis, language processing, and molecule generation, the olfactory-vision model of a robot improves the ability to associate odors with their correct sources, enabling better navigation and decision-making through better sensor selection in critical applications such as explosive detection, drug screening, and search and rescue. It provides a scalable solution to the challenges of limited olfactory data and sensor ambiguity, representing a fundamental advance in the field of artificial olfaction. The code and data are publicly available.
