Embodied AI promises agents that perceive, reason, and act in the physical world, yet realizing such agents demands more than algorithmic novelty. Progress at the frontier increasingly depends on the?co-design?of three tightly coupled layers: the?methods?that learn behavior, the?simulation?infrastructure that supplies the data and experience to learn from, and the?systems?that bridge learned policies to physical hardware. This dissertation argues that the next generation of embodied-AI researchers must be full-stack roboticists, fluent across all three layers, and presents my previous works that together trace this thesis.