Abstract

It is the year of the London Olympics and appropriately, this talk is about making robots run faster, jump higher and throw further. Variable Impedance refers to the ability to change stiffness and damping during task execution. With novel prototype robotic actuators capable of fast impedance modulation, the obvious question is how we can maximally exploit this capability in an automatic and data driven manner? In this talk, I will look at impedance modulation in three different classes of movements: point-to-point tasks like reaching, explosive movement tasks like throwing and rhythmic movement tasks such as walking and running. I will describe an optimal control based formulation of optimizing both the temporal profile of movement and impedance modulation in a way that is tuned to the dynamics of the plant. Several hardware tests will serve to highlight the benefits. Further, I will illustrates the pitfalls of naively mimicking impedance profiles across heterogeneous systems (e.g., human limb to VS joints or MACCEPA actuators) and describe a framework that is capable of abstracting out the specific plant dynamics while ensuring task optimality. This talk will draw upon concepts of optimal feedback control, apprenticeship learning and model free reinforcement learning besides fundamentals of dynamics representation and learning.