Intelligence involves processing sensory experiences into representations useful for prediction. A data-driven approach is proposed to automatically identify state variables for unknown systems from high-dimensional observational data.

The determination of state variables to describe physical systems is a challenging task. Without any prior knowledge of the underlying physics, our algorithm discovers the intrinsic dimension of the observed dynamics and identifies candidate sets of state variables. We demonstrate the effectiveness of this approach using video recordings of a variety of physical dynamical systems, ranging from elastic double pendulums to fire flames. Here we propose a principle for determining how many state variables an observed system is likely to have, and what these variables might be. A longstanding question is whether it is possible to identify state variables from only high-dimensional observational data. Most data-driven methods for modelling physical phenomena still rely on the assumption that the relevant state variables are already known. However, despite the prevalence of computing power and artificial intelligence, the process of identifying the hidden state variables themselves has resisted automation. These variables give a complete and non-redundant description of the relevant system. The push planning method effectively reduces the number of pushes required to move unknown real objects to target positions.Īll physical laws are described as mathematical relationships between state variables. The results show that the prediction model purely trained using our simulation dataset is capable of predicting real object motions accurately. The prediction model and planning method were evaluated both in simulation and real experimental settings. Secondly, we propose a computation efficient planning method that employs a heuristic to reduce the possibility of making sliding contact between the pusher and the object. In this work, firstly, we present a new large planar pushing dataset that contains a wide range of simulated objects and a novel representation for pushing primitives for the data-driven prediction model. However, complex contact conditions and unknown physical properties of the object cause difficulties in reasoning.

To transfer this skill to novel objects, reasoning the pushing effect on object motion is important for selecting proper contact locations and pushing directions. Robot planar pushing is one of the primitive elements of non-prehensile manipulation skills and has been widely studied as an alternative solution to complex manipulation tasks. We also illustrate how its generative nature enables solving other tasks such as outcome prediction. Experiments demonstrate that our model can learn physical properties of objects from video. We propose an unsupervised representation learning model, which explicitly encodes basic physical laws into the structure and use them, with automatically discovered observations from videos, as supervision. Together, they form a dataset, named Physics 101, for studying object-centered physical properties.

We have collected 17,408 video clips containing 101 objects of various materials and appearances (shapes, colors, and sizes). Many physical properties like mass, density, and coefficient of restitution influence the outcome of these scenarios, and our goal is to recover them automatically. We consider various scenarios: objects sliding down an inclined surface and colliding objects attached to a spring objects falling onto various surfaces, etc. Humans can learn basic physical laws when they are very young, which suggests that such tasks may be important goals for computational vision systems. We study the problem of learning physical properties of objects from unlabeled videos.

Physics 101 pdf