Learning Generalizable Robot Policy with Human Demonstration Video as a Prompt