Active Learning and Data Driven Control
We work on machine learning techniques that use control authority (inputs to the system) to actively learn. The inputs do not simply contribute to learning through action, the inputs are chosen to maximize the amount of learning that occurs in a given amount of time. Much of our recent work has used ergodicity as a principle of information coverage. Moreover, we have been using active learning with respect to information about unknown Koopman operators (nonparametric descriptions of dynamical systems).
CPL-Sync: Efficient and guaranteed planar pose graph optimization using the complex number representation
T. Fan, H. Wang, M. Rubenstein, and T. D. Murphey
IEEE Int. Conf. on Intelligent Robots and Systems (IROS), 2019. Winner of ABB Best Student Paper Award. PDF
Local Koopman operators for data-driven control of robotic systems
G. Mamakoukas, M. Castano, X. Tan, and T. D. Murphey
Robotics: Science and Systems Proceedings, 2019. PDF