Keypoint Abstraction using Large Models for Object-Relative Imitation Learning