Chinese robotics startup X Square Robot has open-sourced HOST, a framework that teaches a robot a new manipulation skill from a single human video of about 29 seconds, reporting 62% success where a zero-shot baseline manages 45%. Instead of copying human motions, the robot infers what the task should achieve and works backwards to its own actions; the company says the method needs about one-fiftieth of the data and runs up to 500 times faster than mainstream fine-tuning — without forgetting previously learned skills.