Many vision datasets show what an object looks like; physical AI also needs evidence of how people interact with it and what they intend to do. This release links first-person video, action-level language, and tracked motion, giving models a richer training signal for connecting perception with manipulation.
AI Dataset2026
PhysicalAI Human-1M — 100K hours
Pairs egocentric video with timed task annotations and head-and-hand motion, linking human actions to their intent. The 100,000-hour release supports research and training for physical AI, with a CC BY 4.0 license.
Introduction
Information
- Websitehuggingface.co
- OrganizationsNVIDIA
- Published date2026/11/04
Categories
More Items
Embodied AI2021
Simulates articulated structures interacting with their environment for robotics, biomechanics, and machine learning research. Its combination of generalized-coordinate dynamics and optimization-based contact handling supports control and embodied-AI experiments.
Evaluates whether image-generation systems can retrieve the right evidence from many references, reject distractors, and satisfy entity relations and scene constraints in one image. Expert-written atomic checks pinpoint specific errors.
AI Dataset2026
Unifies metadata for 37,668 3D assets from six libraries, so agents and creators can search by size, style, intended use, and licence. Records link to files or source listings and preserve asset-specific terms.