We present a conceptually and computationally lightweight method for the design and iterative learning of fast maneuvers for quadrocopters. We use first-principles, reduced-order models and we do not require nor make an attempt to follow a specific state trajectory-only the initial and the final states of the vehicle are taken into account. We evaluate the adaptation scheme through experiments on quadrocopters in the ETH Flying Machine Arena that perform multi-flips and other high-performance maneuvers.