Learning from Physical Human Feedback: An Object-Centric One-Shot Adaptation Method