Learning A Physical Long-term Predictor

Abstract

Evolution has resulted in highly developed abilities in many natural intelligences to quickly and accurately predict mechanical phenomena. Humans have successfully developed laws of physics to abstract and model such mechanical phenomena. In the context of artificial intelligence, a recent line of work has focused on estimating physical parameters based on sensory data and use them in physical simulators to make long-term predictions. In contrast, we investigate the effectiveness of a single neural network for end-to-end long-term prediction of mechanical phenomena. Based on extensive evaluation, we demonstrate that such networks can outperform alternate approaches having even access to ground-truth physical simulators, especially when some physical parameters are unobserved or not known a-priori. Further, our network outputs a distribution of outcomes to capture the inherent uncertainty in the data. Our approach demonstrates for the first time the possibility of making actionable long-term predictions from sensor data without requiring to explicitly model the underlying physical laws.

References in corpus (6)

• Very Deep Convolutional Networks for Large-Scale Image Recognition
• Interaction Networks for Learning about Objects, Relations and Physics
• Accelerating Eulerian Fluid Simulation With Convolutional Networks
• A Compositional Object-Based Approach to Learning Physical Dynamics
• Learning to Perform Physics Experiments via Deep Reinforcement Learning
• SMASH: Physics-guided Reconstruction of Collisions from Videos

Cited by in corpus (21)

• Explainable Machine Learning for Scientific Insights and Discoveries
• Deep Learning Methods for Reynolds-Averaged Navier-Stokes Simulations of Airfoil Flows
• Overcoming Limitations of Mixture Density Networks: A Sampling and Fitting Framework for Multimodal Future Prediction
• Physics-Driven Regularization of Deep Neural Networks for Enhanced Engineering Design and Analysis
• Visual Interaction Networks
• Rethinking multiscale cardiac electrophysiology with machine learning and predictive modelling
• gradSim: Differentiable simulation for system identification and visuomotor control
• Comparing recurrent and convolutional neural networks for predicting wave propagation
• IntPhys: A Framework and Benchmark for Visual Intuitive Physics Reasoning
• Approximating the solution to wave propagation using deep neural networks
• Machines Learn to Infer Stellar Parameters Just by Looking at a Large Number of Spectra
• Study on a Poisson's Equation Solver Based On Deep Learning Technique
• Bounce and Learn: Modeling Scene Dynamics with Real-World Bounces
• Occlusion resistant learning of intuitive physics from videos
• Perceiving Physical Equation by Observing Visual Scenarios
• Learning Long-term Visual Dynamics with Region Proposal Interaction Networks
• Learning Generalizable Physical Dynamics of 3D Rigid Objects
• Uncovering Closed-form Governing Equations of Nonlinear Dynamics from Videos
• Predicting the Physical Dynamics of Unseen 3D Objects
• Adding Intuitive Physics to Neural-Symbolic Capsules Using Interaction Networks
• Learning to Transfer Visual Effects from Videos to Images