Skip to content
QwestLabby Kumu
Find tasks
← For contributors

Wearing the camera changes how you move.

Filming from your own point of view feels strange the first time. 5 minutes here will save you re-recording a task — most rejected clips fail for one of the reasons below, and every one of them is easy to avoid once you know it.

Watch first · 4 minutes

How to record a video that gets accepted

Head mount, 0.5× lens, landscape 1080p at 60 fps (30 if your phone can't), both hands in the picture, English narration in the present tense, and ending in the same place you started. Every video must be at least 5 minutes long.

The T-rex pose: a person wearing a phone in a head strap, elbows tucked, both empty hands held up in front of the chest; an inset shows both hands in the camera's picture.
T-rex pose: elbows tucked, empty hands up, both in the picture.

The rule that matters most

Keep both hands in the frame.

This is the single thing reviewers check first, and it is measured automatically. If your hands drop out of shot for a large part of the clip, it cannot be accepted — the footage does not show the robot what it needs to learn.

Overhead first-person view: both hands either side of a box on a desk, with the notebook and plant around it still visible.

Do this

Both hands, work surface visible

Your hands sit in the middle of the frame with the thing you're working on. This is what you're aiming for.

The same desk with the box in the middle of the frame, but both hands cut off below the bottom edge.

Not this

Hands below the frame

Usually means the phone is angled too far up. Tilt it down until you can see your hands while working.

A whole workshop room fills the frame — bench, window, shelving — with the box and the fingertips only just visible along the bottom edge.

Not this

Looking at the room

Natural to do, wrong for this. Look at what your hands are doing, not around the room.

Both hands and the box fill the entire frame, with none of the surrounding desk visible.

Not this

Too close to the surface

Leaning right over the work fills the frame with hands and loses the context around them.

Moving with a camera on

Slow is smooth. Smooth is usable.

A head-mounted camera exaggerates every movement. What feels like a normal glance looks like a whip-pan on playback, and fast motion blurs the exact frames a reviewer needs.

An accepted clip, complete and unedited. Watch how the view moves between the tap, the sink and the rack — following the hands rather than glancing around the room.
  • 01

    Turn your body, not just your head

    Pivot from the waist when you need to look somewhere else. It keeps the picture steady and your hands in shot.

  • 02

    Move about half as fast as feels normal

    Not slow motion — just unhurried. If you can read the labels on things as you pass them, the camera can too.

  • 03

    Let your eyes lead, then your head follows

    Look first, then turn. Whipping your head around to find something is the most common source of unusable footage.

  • 04

    Pause a beat when you change task

    Finishing one thing before starting the next gives reviewers a clean boundary and makes the clip easier to label.

Setting up

Mount it once, check it, then forget it.

Spend two minutes getting the angle right before you start. Adjusting mid-task is the other common reason a clip has to be re-recorded.

An accepted clip, complete and unedited. The camera holds the same view of the work surface for the whole task — that steadiness is what getting the angle right beforehand buys you.
  1. 1

    Strap it on, switch to 0.5×, tilt it down

    The head strap sits just above your eyebrows. In the Camera app pick 0.5× — the wide lens — then tilt the phone down until the sink rim or counter edge is in the top third of the picture, with both hands in the middle. Eye level is too high.

  2. 2

    Record ten seconds and watch it

    Stand where you'll be working and do one normal movement. Check both hands are in frame with the work around them. Adjust the angle down if they aren't.

  3. 3

    Check the light is behind the camera

    A bright window in front of you turns everything into a silhouette. Face away from it, or turn a light on.

The mistake almost everyone makes

Do the task. Don’t demonstrate it.

People slow down, exaggerate, and hold things up to the camera. It is a kind instinct and it ruins the data. Robots need to learn from how the work is really done — the awkward grip, the second attempt, the shortcut you always take.

  • ✕ Holding objects up to the camera✓ Use them the way you normally would
  • ✕ Talking to the camera, or ignoring what the task says about talking✓ Follow the task page. Some tasks want silence. Narrated tasks make talking mandatory: in English, in the present tense ("I am washing the plate", never "I am going to" or "I washed"), the whole way through — live, or as a voice-over added afterwards
  • ✕ Slowing down to be clear✓ Work at your own pace
  • ✕ Restarting when you fumble✓ Keep going — mistakes are useful data
  • ✕ Tidying up first so it looks good✓ Record the mess you'd actually work in

Light

Keep the window behind you, not in front.

A phone exposes for the brightest thing in shot. A head-mounted camera looks where your hands are, so a window beyond them is inside the frame: the phone exposes for the window and your hands fall to a silhouette — the one part that had to be readable. A window behind your head is out of shot and simply lights the work.

The same desk under soft even light: the hands, the box lid and the wood grain are all clearly readable.

Do this

Light falling on your hands

Window or lamp behind you, lighting the work. Even light beats bright light.

The same desk with a bright window directly behind it. The window is blown out white and the hands and box in front of it have fallen into near-silhouette.

Not this

Window in front of you

The commonest lighting rejection. Turn around so the window is behind you.

The same desk in near-darkness. The hands and box are only just distinguishable from the unlit surface around them.

Not this

Too dim to read

If you can't make out the detail on playback, neither can a reviewer. Turn a light on.

One continuous take

Start before you begin. Stop after you finish.

Record the whole task in one go. What a robot has to learn is the order things happen in — reaching before grasping, the pause before the second attempt. Cut those moments out and what is left cannot teach the sequence, however clean each piece looks.

So don’t trim, speed up, filter, or stitch clips together. An edited upload is rejected even when every frame in it would have passed on its own.

Do this

One unbroken recording

Running before the task starts and still running after it ends.

Not this

Separate clips, gaps between them

Each piece may look fine. What happened between them is gone, and that was the data.

Privacy

If it could identify someone, keep it out of frame.

Your footage is reviewed by people and delivered to clients. Anything identifying in it stays in it. This protects you first, and everyone who happens to be nearby second.

Take 30 seconds to clear the surface before you start. It is far quicker than re-recording a task that cannot be accepted.

  • Other people, and children especially

    Nobody but you, including passing through the background. Move to a room you have to yourself.

  • Your own face

    Watch for mirrors, windows at night, dark screens, and anything else that reflects.

  • Screens and notifications

    Phones, laptops, TVs. A message preview appearing mid-task is enough to fail the clip.

  • Documents, IDs, and mail

    Letters, bills, prescriptions, bank cards, anything with a name or an address on it.

  • Anything that locates you

    House numbers, street signs, delivery labels, school uniforms, distinctive shopfronts.

Before you upload

30 seconds that save an hour.

Play back the first and last ten seconds. Almost every avoidable rejection is visible in one of them.

  • Your hands are in frame for most of the clip, and in the middle of it — not along the bottom edge
  • You recorded at 0.5× (the wide lens), unless the task said otherwise
  • Nobody else appears, not even passing through the background
  • Your own face never appears — in a mirror, a window, or a screen
  • No screens, documents, or letters with personal details are readable
  • It's one continuous take with no cuts or pauses
  • The task is finished, not cut short
A real accepted clip: the whole task, start to finish, unedited and unshortened. This is the length and the completeness to aim for — not a highlight of it.

Watch one that passed.

Seeing a single accepted clip end to end teaches more than any checklist. It sets the standard without anyone having to describe it.