Max Planck Institute for Intelligent Systems, Tübingen
Friday, July 19, 2019, 11:00
Room 02-014, Georges-Köhler Alle 103, Freiburg 79110, Germany
Optimal control has been the working horse in modern engineering: from planetary rovers to backflipping robots. Is there anything new that machine learning can bring to the table?
I will start by surveying the main paradigms of learning for control and discuss the current challenges, especially motivated by robot trajectory optimization. Those challenges call for an upgrade of our optimization toolbox to modern robust/stochastic & mixed-integer programming, as well as principled machine learning. Drawing from my experience as an optimizer working with learning tools, I will also introduce a couple of recent projects on learning control that my collaborators and I have been working on at MPI-IS. The talk will end with a discussion on a few interesting recent directions in learning for control.