PHYRE: A New Benchmark for Physical Reasoning

TL;DR

PHYRE introduces a new 2D physics benchmark to evaluate and improve the sample efficiency and generalization of learning algorithms in physical reasoning tasks.

Contribution

The paper presents PHYRE, a novel benchmark for physical reasoning with classical mechanics puzzles designed to foster development of more efficient learning algorithms.

Findings

Modern algorithms perform poorly on PHYRE puzzles

PHYRE encourages development of sample-efficient physics models

Benchmark facilitates evaluation of physical reasoning capabilities

Abstract

Understanding and reasoning about physics is an important ability of intelligent agents. We develop the PHYRE benchmark for physical reasoning that contains a set of simple classical mechanics puzzles in a 2D physical environment. The benchmark is designed to encourage the development of learning algorithms that are sample-efficient and generalize well across puzzles. We test several modern learning algorithms on PHYRE and find that these algorithms fall short in solving the puzzles efficiently. We expect that PHYRE will encourage the development of novel sample-efficient agents that learn efficient but useful models of physics. For code and to play PHYRE for yourself, please visit https://player.phyre.ai.