In the Fall of 2017 I made a two wheel balancing robot in order understand the effect of using a probabilistic extended Kalman filter to compensate for nonlinearities that are present in this inverted-pendulum model. I designed (SolidWorks), laser-cut, assembled, and wired the robot as part of an indivdual project for a course in nonlinear control.
In an extension to this project, regression methods in machine learning (specifically simple linear and support vector regression) were applied to the oscillation pattern of the robot angle in an attempt to predict true angle in near-real time. The intent was to leverage the repetitive nature of the angle readings in balancing.
The results showed that while regression is not yet a viable alternative to traditional filtering methods in control systems, it does track the oscillation and, with further refinement, could be an accurate alternative for data collection in oscillatory dynamic systems. If you are interested, the paper I submitted on this subject can be found here.