Abstract 

1 Reasoning about objects, relations, and physics is central to human intelligence, and 2 a key goal of artificial intelligence. Here we introduce the interaction network, a 3 model which can reason about how objects in complex systems interact, supporting 4 dynamical predictions, as well as inferences about the abstract properties of the 5 system. Our model takes graphs as input, performs objectand relation-centric 6 reasoning in a way that is analogous to a simulation, and is implemented using 7 deep neural networks. We evaluate its ability to reason about several challenging 8 physical domains: n-body problems, rigid-body collision, and non-rigid dynamics. 9 Our results show it can be trained to accurately simulate the physical trajectories of10 dozens of objects over thousands of time steps, estimate abstract quantities such11 as energy, and generalize automatically to systems with different numbers and12 configurations of objects and relations. Our interaction network implementation13 is the first general-purpose, learnable physics engine, and a powerful general14 framework for reasoning about object and relations in a wide variety of complex15 real-world domains.