Abstract

Objects are made of parts, each with distinct geometry, physics,functionality, and affordances. Developing such a distributed, physical,interpretable representation of objects will facilitate intelligent agents tobetter explore and interact with the world. In this paper, we study physicalprimitive decomposition---understanding an object through its components, eachwith physical and geometric attributes. As annotated data for object parts andphysics are rare, we propose a novel formulation that learns physicalprimitives by explaining both an object's appearance and its behaviors inphysical events. Our model performs well on block towers and tools in bothsynthetic and real scenarios; we also demonstrate that visual and physicalobservations often provide complementary signals. We further present ablationand behavioral studies to better understand our model and contrast it withhuman performance.

 

Introduction (beta)

None

 

Conclusion (beta)

None