1: \begin{abstract}

2:   We present a novel approach to estimating physical properties of objects from video.

3:   Our approach consists of a physics engine and a correction estimator.

4:   Starting from the initial observed state, object behavior is simulated forward in time.

5:   Based on the simulated and observed behavior, the correction estimator then determines refined

6:   physical parameters for each object. The method can be iterated for increased precision.

7:   Our approach is generic, as it allows for the use of an arbitrary---not necessarily differentiable---physics engine and correction estimator. For the latter, we evaluate both gradient-free hyperparameter optimization and a deep convolutional neural network.

8:   We demonstrate faster and more robust convergence of the learned method in several simulated 2D scenarios focusing on bin situations.

9: \end{abstract}

10: