Learning beyond simulated physics

Alexis Asseman, Tomasz Kornuta, Ahmet Ozcan

Keywords: double pendulum, chaotic behavior, time-series prediction, real-world physics, spatio-temporal prediction

TL;DR: We introduce a dataset with recordings of the real motion of a double pendulum enabling well scoped research on spatio-temporal predictions of behavior of chaotic systems.

Abstract: Most of recent advancements in spatio-temporal predictions are based on simulated physics. In this paper we introduce a new dataset based on videos of a double pendulum, captured with a high-speed camera, supplemented with positions of its datums and angles between its arms. Because the recorded trajectories depend on unknown initial conditions, the dataset constitutes a benchmark for chaotic behaviors that can be present in other real-world problems. As the description of the system state is extremely simple, the dataset enables careful testing and analysis of the behavior of the developed model. We provide results of stacked LSTM operating directly on angles between arms as a baseline for future research.