X Square Robot open-sources HOST, a one-shot robot skill learner

X Square Robot, a Shenzhen embodied-AI company, announced on 3 August 2026 that it had open-sourced HOST — Human-to-robot One-Shot Skill Acquisition — a method that lets a robot carry out a manipulation task it has never seen after watching one person do it once. The work is joint with Beijing Institute of Technology and Tsinghua University. The preprint went up on arXiv on 22 July (arXiv:2607.20033) and has not been peer reviewed.

The demonstration is a single video averaging about 29 seconds, recording included. Nothing is retrained: HOST updates no model weights at inference time, which is what separates it from fine-tuning methods that need tens of teleoperated demonstrations for each new task. The system locates where the person has got to in the task, predicts what the robot’s own camera should see next, and derives the actions from that prediction rather than from the human’s body, which does not share the robot’s shape.

On novel tasks the authors report 62% average success, measured against baselines fine-tuned on 50 robot demonstrations. They put the data cost at 50 times fewer demonstrations and the acquisition at 507 times faster than supervised fine-tuning.

The code has been public on GitHub under CGuangyan-BIT/HOST since 21 July. The repository carries no root licence file.