A former postdoctoral researcher and engineer at ISIR, Silvia Tulli recently joined the laboratory as a Associate Professor. She looks back on her career, her experiences and the research projects she hopes to develop.
Can you describe your research career path that led you to ISIR?
Silvia Tulli: “My research focuses on learning challenges in robotics. I aim to contribute to physically intelligent systems that collaborate seamlessly with humans. During my PhD, thanks to one of my thesis supervisors, I was introduced to inverse reinforcement learning approaches. I then began to wonder how to integrate information about causal relationships into demonstration learning approaches. The goal was to develop a form of learning that could resemble “why should I do this” rather than simply learning a mapping or reward function that describes the relationship between states and actions. My idea was to use explanation as a learning signal for agents. So I studied both how to generate these explanations, for example via explainable AI, and how to exploit them in robotic learning.
During my postdoctoral research, I expanded this interest to other approaches to interactive robotic learning. I ultimately worked on policy planning in situations where two agents do not share the same model of the task. Subsequently, I worked as a researcher for a company on multimodal machine learning for modelling human biometric responses.”
What struck you most during your first experience at ISIR?
S.T.: “What struck me most were the efforts that some people made to include me, as well as the truly brilliant people I met. The first time I came to ISIR was as part of a European project. I was impressed by the campus (the view from the Zamansky Tower helped!), by the diversity and number of people in the laboratory, and by its central role in the European project and beyond.
Even when I returned later for an exchange programme – a very strange period, as the Covid pandemic began a month and a half after my arrival and continued intermittently for two years – this impression was confirmed. I remember very well that Awatef Barra, even though I spoke hardly any French, helped me understand how things worked.”
You have just been appointed Associate Professor of AI for Robotics (congratulations!). How did you find participating in the competition?
S.T.: “Honestly, I knew the competition was very tough, so I worked hard. During the application period, I was working at a company, which made it difficult to juggle meetings with various people from the university and laboratories and attend the audition during my lunch break.
I received a lot of support, and many people reviewed my project and listened to my presentation for the audition. I also benefited from the valuable insights of several colleagues when I had questions about my professional development.”
What projects do you plan to carry out in the laboratory?
S.T.: “I would like to continue my research on inferring implicit goals when two agents do not share the same task model. I am focusing on sequential decisions using reinforcement learning, imitation learning and deep learning approaches, within the framework of Markov Decision Processes (MDPs) and beyond when the problem requires it. This research is necessary for systems to adapt to disturbances from other agents and their variability.
Applications include multi-agent coordination in robotics, explainable AI to model what the user knows about the system, or, if I may broaden the concept, precision medicine to determine optimal therapeutic sequences based on patients’ genomic profiles.
In all these cases, we have to make assumptions about partially observable models, always favouring evaluation under real conditions, and that is precisely what I find most exciting at the moment.”
Contact : Silvia Tulli, Associate Professor
Published on 01/10/2025
