Two-Stage Learning of Highly Dynamic Motions with Rigid and Articulated Soft Quadrupeds