Learning A Physical Long-term Predictor

TL;DR

This paper presents a neural network approach for end-to-end long-term prediction of mechanical phenomena from sensor data, outperforming traditional physics-based models especially with unobserved parameters, and capturing uncertainty in predictions.

Contribution

It introduces a neural network that predicts long-term physical outcomes directly from sensor data, eliminating the need for explicit physical law modeling.

Findings

Neural networks can outperform physics simulators in long-term predictions.

The approach effectively handles unobserved physical parameters.

It captures uncertainty through probabilistic output distributions.

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.