Humans can easily “see” things beyond visual input: realize common rules in a series of observations and predict if stacked plates will collapse. With our unique ability to abstract and reason, we can quickly discover causality in the environment, master physical rules that the world follows, and interact with anything guided by our minds. Scientific research has shown that even very young infants can summarize and generalize hidden rules from past observations and attempts, utilize prior knowledge for reasoning, and generalize the learned rules to new situations. Towards building machines that can learn and think like people, one natural question for us is whether the intelligence we achieve today manages to understand and adapt to the world, and if any, at what level? Our study focuses on building agents with abstracting and reasoning abilities, including causal discovery, concept learning, intuitive physics, and generalization (e.g., learning tool-use). The ultimate goal is to build agents that can understand the world and explore just like humans.