Research robot learns new skills from one video in 29 seconds

A human performs a task once on video. Seconds later, a two-armed robot tries the same kind of manipulation without a new training run. That is the result reported for HOST, a framework from researchers at Beijing Institute of Technology, X Square Robot and Tsinghua University: an average of 29 seconds to acquire a new skill from one human demonstration.

The system was tested on 50 previously unseen physical tasks, involving different objects, tools and movements. Each task was attempted 20 times, with objects moved to different starting positions or orientations. Human evaluators judged the attempts, producing a 62% average success rate. Phys.org reports that HOST also retained skills the robot had already mastered.

The trick is not to copy the human’s joints. HOST first uses camera images to estimate how far the robot has progressed through the demonstrated procedure. It then predicts the next observations from the robot’s own viewpoint and body structure, before turning those predicted observations into actions. This repeated loop lets the robot follow the human’s procedure while adapting it to its own embodiment.

That matters because conventional skill acquisition can require a costly training-time loop and many task-specific demonstrations. Pandaily describes the prevailing approach as teleoperation followed by fine-tuning, while the researchers report that HOST beat a baseline fine-tuned on 50 robot demonstrations per task and acquired each skill 507 times faster. The framework itself was trained beforehand on 193,462 robot trajectories spanning 229 tasks, then adapted using 5,847 human demonstration videos.

So what changes in practice? A robot operator could record a task once instead of building a large demonstration set for every new variation, potentially shortening the path from a human instruction to a physical trial. But HOST remains a research system: the 62% figure comes from the authors’ evaluation, and Pandaily calls it a benchmark rather than a deployment-ready home robot. The team says it still needs higher success rates and testing on other robots and broader scenarios.