Skip to content
View original post on X: 👩‍💻 Paige Bailey
Published
42/100

Being-H0.5 uses 16,000 hours of human video to scale robot learning

AIBeing-H0.5 researchers present a method that filters egocentric human video and labels it at two timescales for cross-embodiment robot learning. The labels cover 1-second fine-grained motions and 10-second intent summaries, drawing on more than 16,000 hours of footage with Gemini video understanding.
👩‍💻 Paige BaileyVerified on X
@DynamicWebPaige
View on X

Part of a thread · earlier post

Scaling robot learning with 16,000+ hours of human video and Gemini video understanding 🎥

Being-H0.5 filters egocentric footage and generates dual-timescale labels (1s fine-grained motions + 10s intent summaries) for cross-embodiment VLAs.

📍 Oral O1 | Poster S1 #67
🔗 https://openreview.net/forum?id=cfzFrOJmKA

Source details

👩‍💻 Paige Bailey · x.comPublished · added here

Original title: Scaling robot learning with 16,000+ hours of human video and Gemini video understanding 🎥

Topics

AI videoEmbodied AIData & trainingPapers & research