Programming robots that can walk, run and grasp is laborious, so researchers would prefer that they learn on their own. To solve the problem of wear-and-tear on real robots learning by trial-and-error, groups of researchers are developing ways to simulate the bots and download the skills they learn to real hardware. A new method improves these simulations with data from the real robots, closing the feedback loop. The result is robots with boosted speed and agility.