Interconnection and Damping Assignment Control via Reinforcement Learning

Interconnection and Damping Assignment Passivity Based Control (IDA-PBC) is a nonlinear state-feedback control approach which can be used for stabilization and tracking control of a wide range of physical systems. Among many variants of this technique, algebraic IDA-PBC is by far the simplest method as it involves differential-algebraic equations rather than partial-differential equations. However, the equations are generally under-determined and the possible solutions for the unknown elements are non trivial. This issue becomes more evident for tasks on nonlinear systems, for example, swing-up and stabilization of a pendulum or stabilization of electro-mechanical systems. In this paper we mitigate some of the difficulties of algebraic IDA-PBC by using Reinforcement learning. Thanks to the robustness properties of learning the resulting learning based algorithm is insensitive to model uncertainties. We demonstrate the usefulness of the proposed learning algorithm both by simulations and through experimental validation.