Action-Conditional Recurrent Kalman Networks For Forward and Inverse\n Dynamics Learning

Abstract

Estimating accurate forward and inverse dynamics models is a crucial\ncomponent of model-based control for sophisticated robots such as robots driven\nby hydraulics, artificial muscles, or robots dealing with different contact\nsituations. Analytic models to such processes are often unavailable or\ninaccurate due to complex hysteresis effects, unmodelled friction and stiction\nphenomena,and unknown effects during contact situations. A promising approach\nis to obtain spatio-temporal models in a data-driven way using recurrent neural\nnetworks, as they can overcome those issues. However, such models often do not\nmeet accuracy demands sufficiently, degenerate in performance for the required\nhigh sampling frequencies and cannot provide uncertainty estimates. We adopt a\nrecent probabilistic recurrent neural network architecture, called Re-current\nKalman Networks (RKNs), to model learning by conditioning its transition\ndynamics on the control actions. RKNs outperform standard recurrent networks\nsuch as LSTMs on many state estimation tasks. Inspired by Kalman filters, the\nRKN provides an elegant way to achieve action conditioning within its recurrent\ncell by leveraging additive interactions between the current latent state and\nthe action variables. We present two architectures, one for forward model\nlearning and one for inverse model learning. Both architectures significantly\noutperform exist-ing model learning frameworks as well as analytical models in\nterms of prediction performance on a variety of real robot dynamics models.\n