22-DoF dexterous-hand humanoid robot learns fine manipulation from 20,000 hours of human video
Original title:We trained a humanoid with 22-DoF dexterous hands to assemble model cars, operate syringes, sort pok...
AI summary
The research team proposes EgoScale, pretraining GR00T N1.5 on 20,000 hours of first-person human video; with only 4 hours of robot data, it masters highly dexterous tasks such as assembling model cars and operating syringes, improving performance by 54% over training from scratch. The study finds a log-linear scaling relationship between the amount of human video and action prediction loss (R²=0.998). The method leverages the kinematic similarity between 22-DoF hands and human hands, enabling action retargeting without complex transfer algorithms. The policy can transfer across hardware to Unitree G1 (7-DoF), improving performance by more than 30%, and can learn new tasks with only a single demonstration.
Key Takeaways
- The research team proposes EgoScale,pretraining GR00T N1.
- 5 on 20,000 hours of first-person human video
- with only 4 hours of robot data,it masters highly dexterous tasks such as assembling model cars…
- The study finds a log-linear scaling relationship between the amount of human video and action…
- The method leverages the kinematic similarity between 22-DoF hands and human hands,enabling action…
Quick Facts
- Published
- Updated
- Source
- X:Jim Fan (@DrJimFan)
- Category
- Other
人类视频学习呈现完美缩放定律,机器人仅需单演示即可掌握新技能,具身智能迎来数据革命
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: 22-DoF dexterous-hand humanoid robot learns fine manipulation from 20,000 hours of human video. February 26, 2026. https://new.nat6.net/en/news/cmo22oj0y0108slbadann4uen/