HA-Ego-Samples / README.md
humanarchive's picture
Update README.md
2be6f17 verified

Human Archive

Human Archive is modeling human sensorimotor intelligence at scale. We currently collect 2,000+ hours of this multimodal dataset per week, making this the largest and first dataset of its kind.

We’re backed by Y Combinator and engineers from OpenAI, BAIR, SAIL, Anduril Industries, Mercor, NVIDIA, Jane Street, Google, DoorDash AI Research, Reevo, AfterQuery, and the investors behind AMI Labs.

Follow us on X

To purchase the full dataset, find time here

500hr-samples

An egocentric human activity dataset in LeRobot v3 spanning diverse real-world environments, including hotels, restaurants, warehouses, homes, factories, and other commercial and residential settings.

Dataset Overview

Metric Value
Total persons 5,562
Total video segments 10,633
Total duration 499.9 hours
Total size ~2.5 TB
Resolution 1920×1080
Codec H.264
Frame rate 30 FPS
Audio None
Clip duration range 30s – 598s
Unique environments 181
Unique tasks 753

Directory Structure

s3://500hr-samples/
├── commercial/
│   ├── factory/                    # 1,131 persons · 2,013 segments · 106.1 hours
│   │   ├── person{N}/
│   │   │   ├── person{N}_segments.json
│   │   │   ├── person{N}_segment1.mp4
│   │   │   ├── person{N}_segment2.mp4
│   │   │   └── ...
│   │   └── ...
│   └── hospitality/                # 3,619 persons · 7,397 segments · 340.6 hours
│       ├── person{N}/
│       │   ├── person{N}_segments.json
│       │   ├── person{N}_segment1.mp4
│       │   └── ...
│       └── ...
├── residential/                    # 812 persons · 1,223 segments · 53.2 hours
│   ├── person{N}/
│   │   ├── person{N}_segments.json
│   │   ├── person{N}_segment1.mp4
│   │   └── ...
│   └── ...
└── README.md

Split Summary

Split Persons Segments Hours % of Total
commercial/factory 1,131 2,013 106.1 21.2%
commercial/hospitality 3,619 7,397 340.6 68.1%
residential 812 1,223 53.2 10.6%
Total 5,562 10,633 499.9 100%

Per-Person JSON Schema

Each person folder contains a person{N}_segments.json file with the following schema:

{
  "person_id": "person123",
  "total_segments": 3,
  "total_duration_sec": 456.0,
  "segments": [
    {
      "person_id": "person123",
      "video_index": "segment1",
      "duration_sec": 152.0,
      "task": "cutting_vegetables",
      "environment": "commercial_kitchen",
      "width": 1920,
      "height": 1080,
      "fps": 30.0,
      "size_bytes": 245678901,
      "codec": "h264"
    }
  ]
}

Field Descriptions

Field Type Description
person_id string Unique person identifier (matches folder name)
total_segments int Number of video segments for this person
total_duration_sec float Sum of all segment durations in seconds
segments[].video_index string Segment identifier (segment1, segment2, ...)
segments[].duration_sec float Duration in seconds (verified by ffprobe)
segments[].task string Activity label (e.g., washing_dishes, assembling_parts)
segments[].environment string Environment label (e.g., commercial_kitchen, factory_floor)
segments[].width int Video width in pixels
segments[].height int Video height in pixels
segments[].fps float Frames per second
segments[].size_bytes int File size in bytes
segments[].codec string Video codec

Segments Per Person Distribution

Segments Persons %
1 3,028 54.4%
2 1,142 20.5%
3 695 12.5%
4 406 7.3%
5 174 3.1%
6 86 1.5%
7 22 0.4%
8 9 0.2%

Top 30 Environments

Environment Clips Hours
commercial_kitchen 3,679 174.7
factory_floor 1,119 58.6
workshop 1,089 57.7
kitchen 1,036 42.3
warehouse 482 23.4
restaurant 464 23.3
hotel_room 262 9.2
bathroom 249 8.9
office 176 9.0
outdoor_area 171 8.1
home 93 2.6
home_kitchen 85 3.6
cafeteria 83 2.9
bar 75 2.3
banquet_hall 72 3.5
laundry_room 71 4.0
hallway 62 3.2
bedroom 60 2.7
lobby 60 3.2
hotel_lobby 55 2.1
balcony 53 2.1
auto_shop 48 1.8
living_room 44 2.2
restaurant_patio 35 1.5
cafe 35 1.1
conference_room 31 1.3
residential_room 31 1.6
auto_repair_shop 29 0.9
dining_hall 26 1.2

Top 30 Tasks

Task Clips Hours
washing_dishes 793 32.6
wiping_surface 770 31.9
cutting_vegetables 632 33.3
cooking_food 373 16.4
packaging_items 267 11.8
assembling_parts 264 14.1
assembling_footwear 256 13.0
frying_food 236 9.4
portioning_food 234 9.9
operating_machine 180 10.2
mopping_floor 170 8.9
sorting_produce 150 8.0
sweeping_floor 145 6.5
moving_objects 136 4.8
cleaning_floor 134 6.2
folding_clothes 134 4.1
kneading_dough 119 4.4
plating_food 112 5.3
serving_food 106 3.9
washing_mop 103 3.8
arranging_containers 100 4.0
eating_food 95 2.6
scrubbing_surface 94 4.2
preparing_food 92 5.5
clearing_table 87 3.5
packaging_food 86 4.5
polishing_glassware 72 2.4
servicing_machinery 70 2.9
making_bed 65 2.7

Data Quality

All 10,633 video files have been individually verified:

  • Every MP4 passes ffprobe validation (zero corrupt files)
  • No audio streams in any video
  • All duration, resolution, FPS, codec, and file size metadata in JSONs match actual values
  • All segment indices are sequential (segment1, segment2, ...)
  • All total_duration_sec fields match the sum of segment durations
  • No clips shorter than 30 seconds
  • Every person has both a valid JSON and corresponding MP4 file(s)
  • No duplicate videos across the entire bucket

Source

Videos are derived from egocentric recordings captured with chest-mounted cameras at 1920×1080, 30 FPS. Original footage was processed through a QA pipeline using Gemini AI for activity segmentation and quality scoring, then cut into individual segments.