DAGroup-PKU

DAGroup-PKU / HumanNet

Public

HumanNet: Scaling Human-centric Video Learning to One Million Hours

76
0
100% credibility
Found May 12, 2026 at 76 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
AI Summary

HumanNet is an upcoming one-million-hour collection of human-centric videos with annotations for improving AI's understanding of activities, motions, and embodied tasks.

How It Works

1
🔍 Discover HumanNet

You stumble upon HumanNet while reading about cool AI projects that use videos of people doing everyday activities.

2
🌐 Visit the project page

Head to the website to watch demo videos and learn how it gathers a huge collection of human-focused footage from different angles.

3
📚 Dive into the research paper

Read the free paper to understand why a million hours of videos can help AI learn human motions better than robot footage.

4
🔔 Stay tuned for updates

Keep an eye on news for when the full video collection and previews become available to everyone.

5
📥 Grab the dataset

Download samples or the full set of videos with helpful labels about actions, hands, and body movements.

6
👀 Explore the videos

Browse through first-person and third-person clips organized by activities, places, and styles to see real human behaviors.

🚀 Power your AI projects

Use the videos to train smarter AI that understands people and movements, saving time and cost compared to other data.

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 76 to 76 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is HumanNet?

HumanNet curates one million hours of human-centric video, blending first-person and third-person footage with captions, motion annotations, and hand-body signals for fine-grained activity understanding and embodied pretraining. It solves the robot data bottleneck by providing scalable egocentric video as a cost-effective alternative, validated to match 100 hours of real-robot data using just 1,000 hours from HumanNet. Developers access it via Hugging Face datasets with a simple CLI download, organized by multi-axis taxonomy covering viewpoints, tasks, and environments.

Why is it gaining traction?

It stands out with massive scale—one million hours from diverse sources—plus balanced viewpoints and rich annotations that enable motion-aware learning without robot hardware costs. Validation plots show it closing gaps to 20,000-hour robot baselines under vision-language-action protocols, hooking devs needing quick pretraining boosts. Unlike fragmented human network or humanmetrics datasets, HumanNet's curation pipeline ensures quality and privacy.

Who should use this?

Robotics engineers training VLAs on limited real-robot data, like those at humannet chile or humannet iberia scaling humanetics models. Computer vision researchers building motion-centric systems for egocentric tasks, from humannet v3 upgrades to humannet valdivia prototypes. Teams in human metapneumovirus video analysis or humannet viña del mar needing viewpoint-diverse, annotated hours for learning.

Verdict

Promising for human-centric video learning, but at 1.0% credibility, 76 stars, and zero code released yet—it's pre-alpha with just a README and coming-soon dataset. Watch for the Hugging Face drop if you're prototyping embodied AI; skip for production now.

(178 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.