Get a quote
For research labs

Research, on the house.

Up to 100 hours filmed to your brief, free, for a citation.

Applications openuntil , 23:59 IST

Write your research brief
At a glance

The deal.

The research partnership
TermValueWhat it means
Hoursup to 100 hPer lab, split across the workplaces you pick, in steps of 10 h. Plan them in the brief
CostnoneNo fee and no invoice.
Citationone, with the batch idIn every paper, preprint, thesis or release that uses the data, in the format we give. The citation
Data useyour named projectTrain, evaluate and publish results, figures and still frames. No redistribution or re-hosting, no attempt to identify a worker or a site. The data-use summary
Deadline30 November 2026Applications close at the end of that day (IST). Collection follows acceptance, in batches as hours pass review. How it runs
Releaseagreed in writingYour own release (internal use only, a derived dataset or benchmark, model weights trained on it) is agreed before collection starts. Hours collected for you go into our planned open release only if you say so. The open release

A plain-words summary; the binding terms come with the acceptance letter.

See it first

Where you start.

Five things on this site you can inspect before you apply.

1 From the workplaces

Footage

  • LaundryFV-Z00906 · 430.0–438.0 s

  • HVAC servicesFV-C01730 · 600.0–608.0 s

  • MetalFV-B01919 · 620.0–628.0 s

Cut for this page: none of these clips plays anywhere else on the site.

Notes (3)
  • FV-Z00906 · 430.0–438.0 s: camera file 2704×1520 · 59.94 fps
  • FV-C01730 · 600.0–608.0 s: camera file 1920×1080 · 29.97 fps; blurred: the refrigerant cylinder's QR-code label; the same label, lower in the frame.
  • FV-B01919 · 620.0–628.0 s: camera file 1920×1080 · 29.97 fps

Every workplace and its tasks, in the catalogueMore clips play on the home wall

2 RenderRobot hands from our footageEstimated wearer hands, retargeted frame by frame to simulated Allegro and LEAP hands, each final render with its verdict. They show the labels are complete and in sync, not that they are precise.Home, Fig. 3 3 Edited footageRobot video: the arms painted out, a robot drawn inOur re-implementation of Masquerade on our own clips: half meet our rule, both arms edited in every frame. Not camera footage.Home, Fig. 7 4 DataDepth, measuredA lock is fitted and its door measured from two cameras, then rebuilt as a physics model with a simulated hand at its lever. The measurements are repeat readings, never checked against a ruler.Follow the recording 

Data

5 A finding about hands at work

Eight directions hold 91.8% of the posture variance.

55,088 estimated wearer hand poses from four trades: a descriptive finding, not a training result.

The directions, with the hands at work: home, Fig. 4
Share of the posture variance held by each of the 16 directions, largest first

8 of 16 hold 91.8% of the variance

Bars: each direction's share of the variance; line: the running total; dashed: 90%. Each pose is the 16 finger joints of a robot hand, retargeted from one wearer's estimated hand; the directions are the principal components of all of them, from the prior's fit report. The poses are monocular estimates; the detector and wearer rule are not scored against labelled frames.
Who qualifies

For work you'll publish.

Apply if you are

  • A research group at a university or public research institute, with a named principal investigator.
  • Working on a defined research use: a model, a benchmark or a study.
  • A corporate research lab: we decide case by case.

Not for

Your side

One citation, clear rules.

Citation

Use the batch id from each batch's README, so readers can tell which hours you used.

Plain text

Four-Vectors Automation Private Limited (2026). kollected.ai: first-person video of real manual work with inertial data and action labels. Research data partnership, batch <BATCH-ID>, taxonomy v1.0.0. https://kollected.ai

BibTeX

@misc{kollected2026,
  author       = {{Four-Vectors Automation Private Limited}},
  title        = {kollected.ai: first-person video of real manual work with inertial data and action labels},
  year         = {2026},
  howpublished = {\url{https://kollected.ai}},
  note         = {Research data partnership, batch <BATCH-ID>; taxonomy v1.0.0}
}
Data use

You may

  • train and evaluate models on it
  • publish results, figures and still frames from it
  • share it with the co-authors and students of the project you named
  • keep it after the programme ends, for that project

You may not

  • redistribute or re-host the files, in whole or in part
  • try to identify a worker, a colleague or a site
  • use it for a project you did not name, without asking
  • sell it, or build a dataset product from it

We

  • keep the consent record of every batch and show it on request
  • never publish who our partners are, unless you ask us to
Application

Write your brief.

Your answers write the brief; nothing is sent until you press Send.

1 Your lab

2 The research

Distinct workers and sites, for held-out splits
Each worker keeps one pseudonymous id across clips, so you can hold any worker out.
Takes and failures
Release intentRe-release terms are agreed in writing before collection starts.

3 The data

0 of 100 h

Workplaces
All 16 collecting now. Tick a workplace to give it hours; pick its tasks below or leave the shot list to us.
Data types
All available. The programme covers the head camera; the reply says whether it covers the others.
Format (the originals always ship)
Add-ons
Resolution
Conditions of interest

4 Commitments and send

What you agree to

Opens your mail client with a summary of the brief.

We use these details only to answer this application. How we handle your data.

This form needs JavaScript. Email info@kollected.ai with your lab, institution, tasks and hours.

How it works

From brief to batch.

  1. 01 Apply

    Send the brief you build above.

  2. 02 We reply

    By email, with a yes, a question or a no, and the hours we can commit.

  3. 03 Shot list and consent

    Before the first clip, your picks become a shot list for each site.

  4. 04 Collection

    Only clips that pass the acceptance gates enter a delivery.

  5. 05 Delivery

    In batches, by signed link or to your own bucket.

Shot list

Each site's shot list is drawn from the taxonomy, so the footage covers distinct tasks, not one repeated motion.

The taxonomy, as JSON

Consent

Workers sign a written licence for machine-learning training; anyone regularly in shot signs a consent form.

Read the consent policy

Acceptance gates

  • both hands in at least 90% of frames
  • no continuous idle stretch over 10 s
  • an identifiable face in at most 5% of frames
Every published gate

What you receive, per clip

  • FV-B00023/
  • FV-B00023.MP4original camera file, telemetry intact
  • FV-B00023_imu.csvtimestamp_us, ax, ay, az, gx, gy, gz
  • FV-B00023_labels.jsonordered segments with ENV.C.SS codes
  • FV-B00023_meta.jsondevice, mount, site, collector, consent

Formats on request: LeRobot v3, RLDS, HDF5, MCAP. The originals always ship.

Data card

Straight from the files.

Every value is read from the data at build; a dash means unmeasured.

DataFV-B00023 · 450.0–460.0 s

One 10 s clip, four streams, on the video clock

Data only: we publish this clip's motion data, not its picture. Drag across any plot, or focus the time bar and use the arrow keys.

Accelerometerm/s²

ax —ay —az —

Camera axes; gravity is not removed.

Gyroscoperad/s

gx —gy —gz —

Camera axes.

Orientation°

heading —pitch —roll —

Inertial attitude filter (madgwick-imu); heading 0 at the clip's start.

Positionm

x —y —z —

Inertial estimate, not a surveyed path.

Metal workplace: the accelerometer and gyroscope as the camera recorded them, with orientation and position estimated from them.
Inertial data, the same clip: 1,979 samples, 197.92 Hz measured.
ChannelUnitMinMax
axm/s²-5.197-0.211
aym/s²7.47211.252
azm/s²-1.4921.525
gxrad/s-0.4540.424
gyrad/s-0.6220.407
gzrad/s-0.1610.209

The sidecar stores int16 at 1/200 (accelerometer) and 1/500 (gyroscope) of the unit; the delivered CSV is full precision.

Pose, the same clip
FieldValue
OrientationInertial attitude filter (madgwick-imu, β 0.05), gyro bias assumed zero (no still segment in this clip), 5 s pre-roll
Heading drift bound—
Heading · pitch · roll-3.7 to 7.5° · -6.2 to 6.6° · 13.6 to 20.8°
Position sourceimu-estimate
Position noteInertial estimate: gravity-removed acceleration integrated twice with a zero-mean-velocity prior over the clip. Drift-limited; shows sway, not a surveyed path.
Position rangex -0.04 to 0.04 m · y -0.09 to 0.14 m · z -0.02 to 0.03 m

A SLAM trajectory is delivered where the footage supports it; this clip has none.

Clips on this site
ClipWorkplaceFramefpsLengthInertialRate, measuredIMU rowsLabels
FV-B0002315 Metal5312×298859.94588.7 syes197.65 Hz116,5261 · first-pass
FV-B003256 HVAC services1920×108059.94589.4 syes197.36 Hz116,3601 · first-pass
FV-B0191915 Metal1920×108029.97704.7 syes201.13 Hz141,7341 · first-pass
FV-C004569 Carpentry3840×216050442.3 syes200.81 Hz88,9571 · first-pass
FV-Z0090110 Kitchen2704×1520100440.0 syes196.98 Hz86,6732 · first-pass
FV-Z0090210 Kitchen1920×108030120.1 sno——1 · first-pass · narrated
FV-Z0090310 Kitchen1920×108030120.7 sno——1 · first-pass · narrated
FV-Z009067 Laundry2704×152059.94711.7 syes201.10 Hz143,1251 · first-pass
FV-Z0090710 Kitchen2704×152059.94147.6 syes196.24 Hz29,0711 · first-pass
FV-Z0090810 Kitchen2704×152050704.0 syes196.95 Hz138,6532 · first-pass

Most clips carry provisional ids; final ids are assigned at delivery.

An open release, planned.

About 1,000 hours, for anyone to inspect. Not live.

Footage, motion data and labels, planned for Hugging Face. Nothing goes up until its consent review and per-clip approvals are complete.

Who builds the pipeline.

Engineers working with the collection teams on capture, telemetry, labelling and delivery hold affiliations with Georgia Tech, Northwestern University, the University of Pennsylvania, IIT Roorkee, IIT Delhi, IIT Bombay and Carnegie Mellon University. These are individual engineers' affiliations; no institution endorses kollected or reviews its data.

Questions

Before you apply.

Yes, in the builder. Each task we can film gets a new code appended to the taxonomy; existing codes never change.

The 16 in the catalogue, all collecting now.

The catalogue

Still frames and figures, yes. Short clips in supplementary material, usually: ask first, and we check that faces stay under the 5% gate.

Buying for a product, or want to see the files first?

Get a quoteGet free sample data