HOST Paper Reports Robot Skill Learning from One Video in 29 Seconds

Conventional robot imitation systems often require robot demonstrations, task-specific fine-tuning, or repeated training. HOST instead uses a human video as an inference-time guide. Its pipeline estimates how far the demonstrator has progressed through a task, predicts what the robot should observe next in its own embodiment, and then derives actions from those predicted observations.

The authors say this design lets the robot adapt the demonstrated procedure to its own body while leaving the learned policy weights unchanged. That distinction matters because task-by-task fine-tuning can be slow and may degrade previously learned capabilities. In HOST, the demonstration remains a conditioning input rather than becoming a new training set.

Across the paper's evaluation, HOST acquired a new skill in 29 seconds on average and achieved a 62% average success rate. The authors report that it beat their zero-shot baseline by 45 percentage points. They also say it used 50 times fewer demonstrations and acquired skills 507 times faster than a baseline fine-tuned on 50 robot demonstrations per task.

Those figures should be read as paper-reported experimental results, not as proof of reliable general-purpose deployment. The available sources do not establish performance across unrelated robot platforms, uncontrolled environments, or long-running production use. The 62% success rate also shows that one-video acquisition is not equivalent to dependable execution.

For robotics teams, the practical contribution is the inference-time learning pattern: a demonstration can guide a frozen policy without a new optimization cycle. The next evidence to watch is independent replication across different embodiments and tasks, together with failure analysis that explains when the progress-matching cascade breaks down.