Learning Reactive and Predictive Differentiable Controllers for Switching Linear Dynamical Models

Humans leverage the dynamics of the environment and their own bodies to\naccomplish challenging tasks such as grasping an object while walking past it\nor pushing off a wall to turn a corner. Such tasks often involve switching\ndynamics as the robot makes and breaks contact. Learning these dynamics is a\nchallenging problem and prone to model inaccuracies, especially near contact\nregions. In this work, we present a framework for learning composite dynamical\nbehaviors from expert demonstrations. We learn a switching linear dynamical\nmodel with contacts encoded in switching conditions as a close approximation of\nour system dynamics. We then use discrete-time LQR as the differentiable policy\nclass for data-efficient learning of control to develop a control strategy that\noperates over multiple dynamical modes and takes into account discontinuities\ndue to contact. In addition to predicting interactions with the environment,\nour policy effectively reacts to inaccurate predictions such as unanticipated\ncontacts. Through simulation and real world experiments, we demonstrate\ngeneralization of learned behaviors to different scenarios and robustness to\nmodel inaccuracies during execution.\n