Скачать или смотреть 2023 07 14, Bassel El Mabsout, Achieving Robustness in Learned Control

In robot controller design, we try to find controllers which satisfy our high-level intuitions for how they must behave, these notions are sometimes represented as specifications. In this talk, I will be exploring difficulties in defining such controllers when learning is involved. I will then walk through our journey in achieving performant and robust control on real racing quadrotors using Reinforcement Learning. Then I will present an alternative learning framework allowing for finding both learned controllers and certificates where behaviors such as stability are made precise and are proven.

I am Bassel El Mabsout, a Ph.D. student at the Cyber-Physical Systems Lab (http://cpslab.bu.edu/) in Boston University. I’m being advised by Dr. Renato Mancuso and currently work in the domain of robot control. My main research agenda is to build and improve methods for learning real-life robot controllers satisfying behavioral specifications defined by users. In doing so I combine techniques from the domains of Programming Languages, Machine Learning, Control Theory, and Embedded Systems. My contributions mainly entail defining how user intent may be encoded as objective functions in Reinforcement Learning, creating regularization techniques for maintaining performance across domain shifts, and orchestrating systems to perform real world robot actuation.