Skip to content

Dataset · nvidia

Egocentric video and hand poses for physical AI

nvidia/PhysicalAI-Human-1M-100K

An open portion of NVIDIA's Human-1M collection: first-person videos, head and hand poses with text descriptions.

What's inside

The repository features an open subset of 100,000 hours from the Human-1M collection (split size in metadata up to 1,000 records). It contains first-person video recordings, head and hand pose data, images, tabular data, and detailed text annotations in English. The dataset links visual perception, human actions, and intentions.

What you can do with it

  • Training multimodal models for robotics and physical AI
  • Training action recognition and hand/head movement prediction systems
  • Fine-tuning models for generating detailed first-person video descriptions
Languages
English
Size
up to 1K rows
Downloads this month
20.5K
Likes
127

How to load it

from datasets import load_dataset

ds = load_dataset("nvidia/PhysicalAI-Human-1M-100K")