HOST system enables robots to learn skills from a single video in seconds

Summary

The HOST framework, developed by the Beijing Institute of Technology, X SQUARE ROBOT, and Tsinghua University, enables a bimanual robot to acquire new manipulation skills from a single human video. The system learns this skill in approximately 29 seconds of inference time and achieves a 62% success rate across 50 different, unseen tasks without experiencing regression in previously learned skills.

Why it matters

If robots can acquire new manipulation skills from a single everyday human video instead of hours of teleoperation or fine-tuning, the cost and time barriers to deploying general-purpose robots in warehouses, laboratories, and homes are dramatically reduced. The framework’s claim that it preserves existing skills while adding new ones also addresses the long-standing catastrophic forgetting problem in sequential robot learning. This is critical for any robot expected to accumulate an increasing repertoire of skills over its operational lifespan.

Highlights

• 62% average success rate on 50 unseen manipulation tasks; a 45-point advantage over the zero-shot baseline
• Skill acquisition is approximately 507 times faster than a fine-tuned baseline using 50 robot demonstrations per task; uses 50 times fewer demonstrations
• Pre-training on 193,462 robot trajectories (229 tasks) and 5,847 matched human video/robot trajectory examples prior to the 'one-shot' inference phase

Critical take

The 'single video, 29 seconds' headline slightly exaggerates the result. The system has already been trained on nearly 200,000 robot trajectories and 5,800 human videos. Therefore, the one-shot step is a lightweight inference transition added to a heavily pre-trained model, not truly zero-resource learning. Additionally, while a 62% success rate is respectable for a preprint, it means one in three attempts fails. The work is currently an unreviewed arXiv preprint and has been tested on a single bimanual platform. Consequently, the system's adaptability to other robot architectures or more complex, multi-step tasks remains unproven.