Framework for Diversity Auditing and its Application to Datasets in Human-Robot Interaction

The fairness and reliability of AI systems, particularly in Human-Robot Interaction (HRI), are critically dependent on the demographic diversity of their training data. However, many public datasets used to train these systems have significant, often unquantified, biases, which can lead to models that perform poorly or unfairly for underrepresented groups. This Master's thesis proposes to address this challenge by developing a generic, open-source Python tool to audit and quantify the demographic composition (race, gender, age) of visual datasets. Leveraging a pre-trained face analysis model, the tool will provide researchers with a standardized 'diversity report card', enabling a quick and effective assessment of dataset balance before a model is trained. This initial presentation will outline the project's motivation, the detailed research questions, the proposed methodology for the tool's development, and the expected outcomes, including a comparative analysis of several common HRI-related datasets.