Request access to Karigar-100

Please tell us a bit about yourself. We review each request before granting access.

FPV Labs will use the information you provide to review your request, manage access to this dataset and contact you about it. The dataset is released under CC BY-NC 4.0, which does not permit commercial use.

Log in or Sign Up to review the conditions and access this dataset content.

Karigar-100: Egocentric Video of Skilled Work for Robot Learning

Overview

Karigar-100 is an egocentric corpus of real skilled work from FPV Labs, for training and evaluating robot manipulation policies and embodied-AI models. It contains 102.6 hours (1,921 clips) across 10 industries, 14 workplace types, 146 scenes and 90 jobs, recorded at 20 workplaces in 2 countries.

It is weighted towards hard work: 83.3% of the hours are rated Hard, meaning the clip shows sustained, precise hand work such as fitting, alignment and tool control. Every hour was recorded on one of two head-mounted rig types, a six-camera rig or a stereo rig, both with IMU.

Hours 102.6
Clips 1,921 (30 s to 30 min; median 120 s, mean 192 s)
Industries / workplace types / scenes / jobs 10 / 14 / 146 / 90
Workplaces 20, in 2 countries
Difficulty (by hours) 83.3% Hard · 16.7% Medium
Rigs 50: 30 six-camera (6C) and 20 stereo (2C), all with IMU
Size 2.32 TB, one MCAP file per clip
Licence CC BY-NC 4.0 (non-commercial)

What makes it different

  • Mostly hard work. 83.3% of the hours are rated Hard: sustained, precise hand work such as fitting, alignment and tool control. Each sample clip on the dashboard gives the reason for its rating.
  • Real occupational work. Every clip was recorded during a worker's normal job at their own workplace. Examples include brake and wheel service, press brake bending, shoe upper stitching, wood relief carving, carpet weaving and commercial cooking.
  • Multi-view capture. 54.9 h come from six-camera rigs, with camera pairs at the front, side and top of the head, and 47.7 h from stereo rigs. Both rig types record IMU.
  • Labelled and filterable. Every clip carries its industry, workplace type, scene, job, site, rig and difficulty, and download.py fetches just the slice you filter for.

What's in it

Five industries make up 96% of the hours:

  • Industrial manufacturing (28.5 h): sheet-metal bending, drilling, stamping, measuring and marking, and electrical panel wiring.
  • Automotive and transport (19.4 h): brake and wheel service, engine assembly, teardown and repair, and vehicle inspection.
  • Clothing and fashion (18.7 h): shoemaking from pattern cutting and upper stitching to lasting and sole attaching, plus garment sewing.
  • Creative workshops (17.4 h): wood relief and lattice carving, joinery, and carpet loom setup and weaving.
  • Food and beverage (14.1 h): commercial cooking on a kitchen line.

A further 4.4 h covers barbering, farm work, lab sample preparation, hotel housekeeping and print-shop work.

Statistics

Hours by industry

Hours by industry, split into Hard and Medium hours

Hours by workplace type

Hours by workplace type

Top jobs

Hours for the eight jobs with the most footage

All 90 jobs, with a one-line description of each, are on the dashboard.

Clip length

Hours and clips by clip length

Median clip length is 120 s and the mean is 192 s. Half of all clips are between 114 s and 4 min long.

Cameras and format

  • Each clip is one MCAP file, in a folder named after the rig that recorded it (for example FPV-6C-014).
  • 6C rigs (30 rigs, 1,097 clips, 54.9 h) record six cameras: head front left and right, side left and right, and top left and right.
  • 2C rigs (20 rigs, 824 clips, 47.7 h) record a stereo pair: top left and right.
  • Both rig types record IMU alongside the video.

Labels

Each clip is labelled with:

Label Description Filter
Industry L1, e.g. Food and Beverage --industry
Category L2 workplace type, e.g. Auto Repair --category
Scene L3 scene within the workplace, e.g. Workbench --scene
Job One of 90 jobs, e.g. Metal welding --job
Site Workplace the clip was recorded at --site
Rig 6C (six cameras) or 2C (stereo) --rig
Difficulty Hard or Medium, from a per-clip rating --difficulty

Download

Each clip is one MCAP file at episodes/<device>/<episode_id>.mcap. manifest.csv lists every clip with its labels, size and checksum. download.py filters it and downloads only the matching files.

pip install -U huggingface_hub
hf auth login
hf download fpvlabs/karigar-100 download.py manifest.csv --repo-type dataset --local-dir .
Filter Example
--industry "Food and Beverage"
--category "Auto Repair"
--job "Metal welding"
--scene Workbench
--site "Commercial Kitchen"
--rig 6C (6 cameras) or 2C (stereo)
--difficulty Hard or Medium
--max-hours 2 (random sample of the matches)

Use the names exactly as the dashboard shows them, or part of a name (food, welding); case is ignored. Repeat a flag to match any of its values; different flags must all match.

python download.py --list job                               # all jobs with clips / hours / GB
python download.py --industry clothing --rig 6C --dry-run   # size of the selection, nothing downloaded
python download.py --job "Metal welding" --difficulty Hard
python download.py --category "auto repair" --max-hours 2

Full lists of industries, categories, jobs, scenes and sites are on the dashboard or from python download.py --list <field>.

How it was made

  • Recording: workers at each site wore head-mounted camera rigs during their normal work.
  • Labels: each clip is labelled with its environment (industry, workplace type and scene), job and difficulty.
  • Difficulty: a per-clip rating of how much sustained precise work the clip shows, such as fitting, alignment and tool control. The dashboard's sample clips show the reason given for each rating.

Intended uses

  • Pre-training and fine-tuning manipulation and embodied-AI policies.
  • Stereo and multi-view learning from head-mounted cameras and IMU.
  • Video understanding of skilled work: action, tool use and task recognition.
  • Evaluating models on hard, precise manipulation.

Licence

This dataset is released under CC BY-NC 4.0. You may use it for research and other non-commercial work, with attribution to FPV Labs.

Commercial use is not permitted. This includes training or evaluating models for commercial products or services. For a commercial licence, contact FPV Labs at contact@fpvlabs.ai.

Contact

FPV Labs · fpvlabs.ai · contact@fpvlabs.ai

Downloads last month
728