ac-RKN

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

  • Author:

    Vaisakh Shaj, Philipp Becker, Dieter Buchler, H Pandya, Niels Duijkeren, J Taylor, M Hanheide, Gerhard Neumann

  • Source:

    Conference on Robot Learning (CoRL), 2020

  • Date: 19-10-2020
  • We adopt a recent probabilistic recurrent neural network architecture, called Recurrent Kalman Networks (RKNs), to model learning by conditioning its transition dynamics on the control actions. Inspired by Kalman filters, the RKN provides an elegant way to achieve action conditioning within its recurrent cell by leveraging additive interactions between the current latent state and the action variables. We present two architectures, one for forward model learning and one for inverse model learning that leverage the disentagled representation of state and action variables in the latent space. Both architectures significantly outperform existing model learning frameworks as well as analytical models in terms of prediction performance on a variety of real robot dynamics models.

    Source Code: https://github.com/ALRhub/action-conditional-rkn

    Preprint: https://arxiv.org/pdf/2010.10201.pdf