뉴스

D-Report: "Passing Through a Needle's Eye with Ease"... Reducing Data, Increasing Success Rates

A robot passes a thread through a narrow needle's eye.

It moves effortlessly without colliding with obstacles, even when there is only 2.5mm of clearance.

The robot performs delicate and precise movements—tasks that would require intense concentration from a human—without a hitch.

While such precise movements previously required vast amounts of data and training, researchers at KAIST have developed a technology that enables learning with only a small amount of data.

By training a model that learns changes over time alongside a model that generates various actions, the research team developed a robot that shows better performance even when trained on only one-quarter of the data used in previous methods.

The research team explained that while previous studies treated every moment of human movement with equal weight, this study placed less emphasis on large movements and focused heavily on learning precise movements, such as manipulating a thread.

The robot's performance also showed a significant difference.

Experimental results showed a task success rate up to 81% higher than that of existing robots.

[Interview: Park Dae-young / Professor, School of Computing, KAIST: "Because we now take dynamics into account, the robot understands and learns the difference between fast and slow. Since it demonstrated high performance in areas requiring very precise control, we expect it to be applied to areas in our country that are not yet automated."]

As the technology saves on the data required for training, it can also address issues related to GPU and memory shortages.

The research team expects this technology to be applied to various industrial fields that require high accuracy, such as precision parts assembly, manufacturing, and medical surgery.

Reported by Seo Donggyun | Video by Choi Hye-young | Produced by SBS Digital News
※ Please note: This article was translated by AI and may contain errors.
Copyright Ⓒ SBS & SBSi. All rights reserved.
Copying, redistribution, and unauthorized use in AI training are strictly prohibited.

Most Read