Open Research Dataset • Actively Collecting

HumanLoop Egocentric Dataset

First-person recordings of everyday household activities, captured with head-mounted and body-worn cameras and annotated with hand pose, segmentation, and audio narration — built for hand-pose, manipulation, and VLA / robotics research.

Footage Captured 50+ Hours

What is HumanLoop?

HumanLoop is a first-person dataset of everyday household activities captured in native home environments. Using head-mounted and body-worn cameras, we record unscripted interactions across the entire home — kitchen, laundry, living spaces — annotated with hand pose, segmentation, and action labels.

Example Activities

washing a cup
cutting banana
arranging shoes
folding clothes
washing clothes
making tea
mopping
sweeping
pouring water
coffee from cupboard
milk from fridge
milk in fridge
organising clothes
throwing trash
washing a whisk

Data Quality Gates

Every recording is scored 0–100 by pipeline/quality_check.py before it's added to the dataset.

Check Threshold Score Weight
Frame Rate ≥ 29.9 FPS (GoPro standard) +10 pts
Resolution Height ≥ 1080p +5 pts
Hand Detection Rate ≥ 80% of sampled frames (MediaPipe Hands) +25 / −20 pts
Both Hands Visible ≥ 50% of sampled frames +15 pts
Sharpness (Blur Score) Laplacian variance ≥ 500 +20 / −25 pts
Brightness Avg pixel value between 80–180 +15 pts
Audio Narration Speech detected via Whisper (no_speech_prob < 0.5) +10 / −15 pts