Abstract

Numerous demonstrations have highlighted the potential of learning-based control paradigms over the last years. A number of both practical and theoretical challenges arising from the use of learning in closed-loop control systems, however, are still limiting the widespread success of these promising techniques in practice. In this talk, I will focus on the questions of data-efficiency and safety. In order to provide a scalable and flexible approach for complex systems, the techniques will build on optimization-based control. The first part of the talk will discuss techniques for learning objective and dynamics models, which are both real-time capable and allow for fast online adaptation despite limited initial data and information for a particular task at hand. In the second part, I will present the concept of a predictive safety filter to augment any learning-based controller with safety guarantees and will outline connections to techniques based on control barrier functions. The results will be highlighted using examples from robotics.

The seminar will be held in the JDB Seminar Room, Department of Engineering, and online (zoom): https://newnham.zoom.us/j/92544958528?pwd=YS9PcGRnbXBOcStBdStNb3E0SHN1UT09