Intelligence is the ability to do the right thing in myriad unfamiliar situations. Robots today can do very well in complex, but narrow tasks, e.g., in-hand manipulation. However, home service robots (if any) are far from their human counterparts in abilities. What is the nature of the underlying gap? How would the latest advances in AI -- learning, planning, foundation models -- help to overcome this gap? In this talk, I will argue that integrating model-based planning and data-driven learning will lead to a data-driven, compositional learning architecture for scalable robot intelligence. The key issue here is the interplay, rather than the conflict, between structure and data. I will illustrate the general thinking with our work on robots navigating anywhere on our university campus, robots folding a variety of clothes, and robots aiming to operate in the open world.