Retargeted quadruped motions from monocular videos via a learnable skeleton, removing the need for motion capture or manual keypoints.