Hoppa till huvudinnehållet

Ioanna Mitsioni

Profilbild av Ioanna Mitsioni

DOKTORAND

Detaljer

Adress
LINDSTEDTSVÄGEN 24

Om mig

I have completed my PhD in the division of Robotics, Perception and Learning at KTH, under the supervision of Danica Kragic and the co-supervision of Yiannis Karayiannidis.

I received my  Diploma of Electrical and Computer Engineering from the Aristotle University of Thessaloniki, where I also did an internship in the Automation and Robotics Lab on the topic of Machine Learning methods for slippage detection (more information about this project can be found on this page).

My research is mainly in the context of  contact-rich manipulation tasks and I am investigating how we can address safety aspects that occur when we employ data-driven solutions to encompass the complicated dynamics of those tasks.

More specifically, to safely deploy data-driven approaches on physical systems:

  • We need good predictive performance, since we lose the ability to predict the exact responses from an analytic model.

  • We need the controller to guide the system to safe behaviors or otherwise abort the tasks, even in the presence of imperfect predictions.

  • We need to understand the models we use, as most networks are black-boxes which makes evaluations difficult.

To address the first point, I have worked on modelling the dynamics of food-cutting for a velocity-resolved data-driven controller. In addition to this, I have investigated how adjusting the baseline controller used for data collection and using a curriculum while training, can produce models with better performance and understanding of the long-term dynamics (more information about these works can be found on this page).

To address the second point,  I have worked on a method that learns a set of safe states through positive and negative examples and avoids control inputs that may lead the system to unsafe states.  This approach led to my latest work on safe data-driven control that does not require negative examples and actively leads the system towards safe states, reducing the risk for the system even further (more information about these works can be found on this page).

Finally, to address the last point I have worked on the evaluation of learned models based on interpretability modules that allow us to understand what led networks to make a decision (more information about this work can be found on this page).

A general overview of my work can be found in my 80% seminar "Safety Aspects of Data-Driven Control in Contact-Rich Manipulation" (October 20th 2021) below.

The slides for this seminar can be found here.