Learning to control

Team:

NameDepartment
Shishir N. Y. KolathayaRBCCPS
Bharadwaj AmruturRBCCPS/Electrical Communication Engineering (ECE)
Shalabh BhatnagarComputer Science and Automation (CSA)

We are interesting in the fusion of learning and control in areas like:

  1. Manipulation
  2. Vehicle navigation
  3. Legged locomotion

The field of “learning to control” is getting a lot of attention of late. Traditional methods of applying control involve a rigorous formal analysis of stability, convergence rates, and also stability margins for a wide variety of nonlinear dynamical systems. But these methods often become very complex and computationally expensive as system complexities increase. On the other hand, learning based methods have a huge advantage of using prior data to obtain seemingly simple control tasks for realizing a wide variety of complex behaviors in nonlinear systems. But these methods do not provide formal guarantees of safety and stability for any of the learnt tasks obtained.

Therefore our objective is to explore control methodologies that take the best from both of these two approaches: Obtain control laws that not only guarantee sufficient performance and stability margins, but also are fast and simple to implement.

Back