Quadcopters are among the most agile and dynamic machines ever created. In this talk, I’ll show how data-driven sensorimotor controllers can push quadcopters with only onboard sensing and computation to their physical limits. Such controllers enable quadcopters to fly faster and more agile than what was possible before in unstructured environments like cities, forests, and disaster scenarios. The insights acquired from quadcopter flight transfer to other domains, including legged locomotion and bio-inspired vision. However, fundamental research questions still need to be addressed to make agile robots adaptable, robust, and safe, and enable their wider application in homes, search and surveillance, and inspection. Teaser videos of the results I will present are Deep Drone Acrobatics (https://youtu.be/2N_wKXQ6MXA) and Agile Autonomy (https://youtu.be/m89bNn6RFoQ).