The project HUmLrn – Unified Learning from Diverse Human Feedback, presented by Prof. Alberto Metelli from the Department of Electronics, Information and Bioengineering – Politecnico di Milano, has been selected as one of the winners of the FIS 2 – Italian Science Fund call.

HUmLrn tackles one of the most pressing challenges in artificial intelligence: how to effectively leverage the diversity of human feedback to enhance the capabilities of artificial agents. When an expert interacts with a system, their behaviour conveys implicit information about the goals they are trying to achieve, offering valuable insights for machine learning.

Traditional methods rely on a narrow range of feedback types—typically limited to demonstrations—or assume the availability of a predefined reward function. As a result, they overlook the wide variety of signals that humans can provide, such as corrections, preferences, or guidance, and fail to consider the potential of integrating them in a coherent and effective manner. In contrast, humans naturally combine diverse forms of feedback during learning, extracting value from each and adapting their behaviour accordingly.

HUmLrn proposes a fundamentally new approach, aiming to build a unified framework capable of managing and leveraging heterogeneous forms of human feedback. Starting from a solid theoretical foundation, with particular focus on the statistical complexities involved, the project aims to design and develop innovative and efficient learning algorithms that can handle the richness and variability of the input they receive. The project will culminate in an experimental validation phase, through a simulated use case in the field of autonomous driving, where the learned capabilities will be tested in a realistic environment.

With HUmLrn, the Politecnico di Milano positions itself at the forefront of research in machine learning, promoting a richer, more flexible, and collaborative vision of the interaction between humans and artificial intelligence.