Replay Overshooting: Learning Stochastic Latent Dynamics with the Extended Kalman Filter

IEEE International Conference on Robotics and Automation (ICRA) 2021

Title: Replay Overshooting: Learning Stochastic Latent Dynamics with the Extended Kalman Filter

Authors: Albert H. Li*, Philipp Wu*, Monroe Kennedy III

(*=equal contribution)

Abstract:

This paper presents replay overshooting (RO), an algorithm that uses properties of the extended Kalman filter (EKF) to learn nonlinear stochastic latent dynamics models suitable for long-horizon prediction. We build upon overshooting methods used to train other prediction models and recover a novel variational learning objective. Further, we use RO to extend another objective that acts as a surrogate for the true log-likelihood, and show that this objective empirically yields better models than the variational one. We evaluate RO on two tasks: prediction of synthetic video frames of a swinging

motorized pendulum and prediction of the planar position of various objects being pushed by a real manipulator (MIT Push Dataset). Our model outperforms several other prediction models on both quantitative and qualitative metrics.