From Theory to Practice: Overcoming the Real-World Challenges of Reinforcement Learning
PI Story with Marcello Restelli
The Doctoral Programme in Information Engineering and Computer Science (IECS) organizes the sixth edition of the PI Stories, a series of seminars aimed at providing the opportunity for PhD students to learn the success stories of some of the most talented researchers in the world.
Each speaker will present a research project he/she led as a principal investigator. The presentation will cover the scientific scope of the project and the most important results that the project has achieved. The speakers will also share their own experiences of turning a research idea into a successful project winning a competitive grant.
The events will take place in presence or online on Zoom and will be held in English.
Abstract
Reinforcement learning (RL) has achieved remarkable progress in simulated environments, yet deploying RL in real-world systems remains profoundly challenging.
This lecture will explore the key obstacles that arise when moving from controlled benchmarks to dynamic, high-stakes applications such as robotics, healthcare, and industrial decision-making. We will examine issues related to sample inefficiency, safety and risk sensitivity, reward design, distributional shift, partial observability, and the difficulty of integrating prior knowledge. The talk will also highlight practical concerns, including unreliable simulators, expensive data collection, non-stationary environments, hardware constraints, and the need for human oversight. Finally, we will discuss emerging solutions - such as offline and batch RL, model-based approaches, safe RL frameworks, and hybrid learning-control architectures - that aim to bridge the gap between theory and deployment. Attendees will gain a clear understanding of why real-world RL is hard, what progress is being made, and where impactful research opportunities remain.
Bio
Marcello Restelli is a Full Professor at the Department of Electronics, Information and Bioengineering at Politecnico di Milano, where he coordinates the Real-Life Reinforcement Learning Research Lab (RL3). He is the author of more than 200 international scientific publications, primarily focused on the study and development of new reinforcement learning techniques. His research results are applied to real-world problems through numerous industrial collaborations in diverse sectors, including finance, e-commerce, Industry 4.0, and automotive. He is an ELLIS Fellow and serves as the research lead for the Artificial Intelligence Observatory of Politecnico di Milano.
In 2020 and 2024, he co-founded ML cube and Trade RL, two spin-offs of Politecnico di Milano, where he is currently scientific advisor.