Learning A Physical Long-term Predictor

By Sébastien Ehrhardt, Áron Monszpart, Niloy J. Mitra, Andrea Vedaldi. 2017

67 citations recorded by the source corpus.

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.

Subfields: Artificial Intelligence, Signal Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Signal Processing

Citations in the corpus

Latent Space Physics: Towards Learning the Temporal Evolution of Fluid Flow (2018, background): “predict two-dimensional rigid bodies physics [Chang et al. 2016]. More recently, Ehrhardt et al. use recurrent NNs to predict trajectories and their likelihood for balls in height-field environments [Ehrhardt et al. 2017]. Others targeted predictions of liquid motions for robotic control, with a par”
Visual Interaction Networks: Learning a Physics Simulator from Video (2017, methodology): “A third class of methods [5, 8, 9, 23], like our Visual Interaction Network, have been used to predict future state descriptions from pixels.”
Unsupervised Discovery of Parts, Structure, and Dynamics (2019, background): “Physical scene understanding has attracted increasing attention in recent years (Fragkiadaki et al., 2016; Battaglia et al., 2016; Chang et al., 2017; Finn et al., 2016; Ehrhardt et al., 2017; Shao et al., 2014).”
Taking Visual Motion Prediction To New Heightfields (2017, unknown): “Ehrhardt et al. (2017b) showed, how long-term extrapolation models can be trained for one object moving on smooth analytic surfaces, such as ellipsoids.”
Explainable Machine Learning for Scientific Insights and Discoveries (2019, background): “Further details, types of interpretation, and specific realization can be found in recent surveys [7], [8], [20].”
Generating Liquid Simulations with Deformation-aware Neural Networks (2017, background): “Others have targeted this goal in specific settings such as robotic interactions [14], sliding and colliding objects [43, 42], billiard [15], or trajectories in height fields [12].”
IntPhys: A Benchmark for Visual Intuitive Physics Reasoning (2019, background): “Results are detailed in Supplementary Materials (Tables 11, 12, 13, 14).”
IntPhys: A Framework and Benchmark for Visual Intuitive Physics Reasoning (2018, background): “In [21] and [17], authors use a visual decoder to reconstruct the future frames, allowing the model to learn from raw (though synthetic) videos.”