Learning to Represent Mechanics via Long-term Extrapolation and Interpolation

Abstract

While the basic laws of Newtonian mechanics are well understood, explaining a physical scenario still requires manually modeling the problem with suitable equations and associated parameters. In order to adopt such models for artificial intelligence, researchers have handcrafted the relevant states, and then used neural networks to learn the state transitions using simulation runs as training data. Unfortunately, such approaches can be unsuitable for modeling complex real-world scenarios, where manually authoring relevant state spaces tend to be challenging. In this work, we investigate if neural networks can implicitly learn physical states of real-world mechanical processes only based on visual data, and thus enable long-term physical extrapolation. We develop a recurrent neural network architecture for this task and also characterize resultant uncertainties in the form of evolving variance estimates. We evaluate our setup to extrapolate motion of a rolling ball on bowl of varying shape and orientation using only images as input, and report competitive results with approaches that assume access to internal physics models and parameters.

arXiv admin note: text overlap with arXiv:1703.00247

References in corpus (5)

• Visual Dynamics: Probabilistic Future Frame Synthesis via Cross Convolutional Networks
• Learning Visual Predictive Models of Physics for Playing Billiards
• Learning to Perform Physics Experiments via Deep Reinforcement Learning
• Predicting Deeper into the Future of Semantic Segmentation
• Visual Stability Prediction and Its Application to Manipulation

Cited by in corpus (3)

• Bounce and Learn: Modeling Scene Dynamics with Real-World Bounces
• Learning Generalizable Physical Dynamics of 3D Rigid Objects
• Predicting the Physical Dynamics of Unseen 3D Objects