Grape Preprint

New preprint: GRAPE, a trajectory-wise preference alignment method for post-training VLA models.