The world embraces various forms of structure in it. It is governed by time and space (3D), partial-observability, uncertainty, causality, and also is composed of modular entities like objects. The power of human intelligence seems to come from the ability to discover these structures and utilize those in such a way to support systematic generalization. Furthermore, a human-level AI would need to achieve this ability in an unsupervised fashion like humans. In this talk, I introduce recent research efforts of my research group toward this direction, specifically via unsupervised object-centric representation learning.