Startup Zizhu Robot has open-sourced HOST, a framework that aims to collapse the cost of teaching a robot a new skill by letting it learn from a single short human demonstration video instead of repeated data collection and retraining.
Learning the outcome, not the motion
HOST, for Human-to-robot One-Shot Skill Acquisition, has the robot watch an average 29-second human video and reproduce a new skill with 62 per cent success. The key design choice is to ignore the exact human movement and instead understand the task’s outcome, the change in the world, then plan the robot’s own motion to reach it. A person may grip a cup with five fingers; a robot may use a two-finger gripper. As long as the target is ‘lift the cup steadily’, each body finds its own path, and the visual result becomes the shared language between human and machine.
Align by progress, not by clock
Because humans and robots move at different speeds, HOST aligns by task progress rather than elapsed time, mapping each video frame and each robot step into one vector space to find the matching stage. Zizhu says this cuts temporal-alignment error by an order of magnitude. A dual-expert architecture splits the work: a ‘video expert’ reads the scene and predicts the next result, and an ‘action expert’ converts that result into motion. After pretraining the weights are frozen, so new skills are acquired at inference time.
Why it matters
Against Pi-0.5, which needs about 50 robot demonstrations and roughly four hours of fine-tuning for 38 per cent success, HOST uses one-fiftieth of the data and one-five-hundredth of the time at a higher success rate, while older video-imitation methods sit near 19 per cent. Just as important, fine-tuning rewrites model weights and retains only 17 per cent of prior skills, whereas HOST preserves what it already knew. At 62 per cent the method is far from solved, but it shows robot skill acquisition need not mean endlessly rewriting the model.
Read the original report (OFweek Robotics)
Translated and adapted from OFweek Robotics (robot.ofweek.com).