On March 12th, the session Scalable and reliable collaborative autonomy via online learning, coordination, and control will take place, featuring Professor Vasileios Tzoumas.

Vasileios Tzoumas is an assistant professor at the University of Michigan, where he leads the Intelligent Robotics and Autonomy Lab (iRaL), and chairs the Academic Success, Community, and Outreach Committee. Previously, Vasileios was at MIT as a research scientist, and at UPenn as a Ph.D. student.  His research is on scalable and reliable cyber-physical systems in resource-constrained, unstructured, and contested environments, such as the systems found in defense, disaster response, and smart cities.  He works at the intersection of control, robotics, and online optimization and learning.  Vasileios is the recipient of an NSF CAREER Award, the Best Paper Award in Robot Vision at the 2020 IEEE International Conference on Robotics and Automation (ICRA), an Honorable Mention from the 2020 IEEE Robotics and Automation Letters (RA-L), and a Best Student Paper Finalist Award at the 2017 IEEE Conference in Decision and Control (CDC).

Abstract:
Roboticists envision autonomous robot systems that collaboratively complete safety-critical tasks ranging from mapping to inspection to target tracking.  Such tasks, however, are currently challenging to complete rapidly and reliably.  One reason is that they often take the form of NP-hard combinatorial optimization problems.  Another reason is that they can necessitate operation in resource-constrained, unstructured, and unknown environments.  For example, the robots may have access only to slow vehicle-to-vehicle communication, can be challenged by mobile obstacles with unknown dynamics, and can be exposed to difficult-to-model disturbances such as ground effects, turbulence, and wind.  Current near-optimal approaches cannot scale to more than a handful of robots and rely on conservative assumptions about the unknowns compromising accurate control.

I will present near-optimal coordination and control methods to overcome the above challenges.  On coordination, I will present near-optimal algorithms that scale to tens of robots.  Critical to our approach is a novel decentralized optimization framework that enables the suppression of communications while accounting for the impact on the approximation performance.   On control, I will present model predictive control methods that adapt to the unknown dynamics and disturbances at hand, guaranteeing no regret.  Critical to our approach is a self-supervised online system identification method for non-linear and state-dependent disturbances.  We use tools from nonlinear control, submodular optimization, and machine learning.  I will demonstrate the methods in simulator and hardware experiments.  I will end the talk advocating the need for the online co-adaptation of hardware, algorithms, and their networks for scalability and reliability.