Can Machine Learning Understand and Predict Physics?

Title:

Can machine learning understand and predict physics? A brief review on state of the art “Intuitive Physics” and its potential application on robotics and medicine.  

Abstract:

Modelling the physical properties of objects is a fundamental prerequisite for understanding the environment and interacting with it. An emerging area of research tackles how to learn intuitive physics with artificial neural networks, in order to infuse deep networks with the ability to predict the future.

In this work we propose the use of conditional generative adversarial networks for predicting body deformations under the effect of external forces from a single RGB-D image. We propose a new way to encode the physical properties of the material and the applied force, and we compare it with a baseline. The network is based on an Invertible conditional Generative Adversarial Network (IcGAN) and is trained on a collection of different objects of interest generated by a finite element model simulator. The network is able to reconstruct the whole 3-D structure of the object given a single depth view and to generalise to unseen objects. The encoding of the physical properties of the material improves generalisation over material types, relaxing the need for a large training set of different body materials. Our approach is fast enough to be used in near real-time applications, contrary to traditional finite element simulations, and offers more generalisation abilities compared to previous learning methods. We apply the network to the problem of navigation of mobile robots carrying payloads over different floor structures and materials.

Bio:

Zhihua Wang, third year DPhil student supervised by Prof Andrew Markham, advised by Prof Niki Trigoni in Department of Computer Science. Zhihua’s research interests include physics modelling, machine learning, sensor networks and magnetic sensing.