Abstract:
Safety and robustness in the context of cyber-physical systems (CPS) have always been critical, particularly when there is close interaction with human users. Over recent years however, certifying such properties for machine learning algorithms has also risen in importance, given the increasing use and close integration of such algorithms in CPS for tasks such as perception, prediction, and even decision making. Viewed under the lens of dynamical systems, the behavior of these disparate components of a CPS may be studied under a control-theoretic framework. Having a common set of tools to analyze these various systems, physical or otherwise, can be convenient because oftentimes, they work together as sub-systems of a larger, interconnected system and individually tuning each component can lead to a conservative design.
In my talk, I will present some of my analytic results for robotic and learning systems, focusing on safety, stability, and robustness problems in a control-theoretic setting, and touch upon some of my recent data-driven techniques based on Koopman operators, for learning in control. In particular, I will highlight my work on Lyapunov and differential games inspired design of controllers for safe bilateral teleoperation of surgical robots that are robust under time-delays and input disturbances, as well as my work on real-time adversarial attacks on Recurrent Neural Network classifiers. Finally, I will envision them coalescing in the domain of robot assisted tele-health as one of the direction that I plan to pursue in the future.
Biosketch:
Shankar Deka is currently a postdoctoral researcher in the Hybrid Systems Lab supervised by Claire J. Tomlin in EECS department at University of California Berkeley, where he works in problems at the intersection of nonlinear control theory, machine learning, and robotics. Prior to joining Berkeley, Shankar was in the Coordinated Science Laboratory at the University of Illinois, Urbana-Champaign, where he obtained his PhD and MS degrees in Mechanical Engineering in 2019 and 2016 respectively. Advised by Dušan M. Stipanović and Thenkurussi Kesavadas, his doctoral research focused primarily on the control of multi-agent Lagrangian systems in context of medical robotics. He obtained his BS degree in Mechanical Engineering from Birla Institute of Technology and Science, Pilani in 2014.
