The dataset
0 accepted demonstrations, 0.0 minutes of hand motion from 0 people. Every clip comes with 21 tracked points per hand for every frame.
Take it with you
The manifest and landmark files update live. Snapshots are built once a day and include the clips.
Accepted demonstrations
Hand motion by default. Switch a card to the clip to see the video.
About this data
What is in it
Short videos of human hands doing everyday tasks, each about 10 seconds long, recorded on phones and laptops.
Every clip comes with per-frame hand landmarks: 21 points per hand in image space and in metres (world space), which hand it is (left or right), a confidence score and a timestamp in milliseconds.
Clips are re-encoded to H.264 MP4 at up to 720p and 30 fps, with no audio and no file metadata.
How it was collected
Contributors pick a task on mimicrobots.fun and record themselves doing it. Hand tracking (MediaPipe Hand Landmarker) runs on the contributor's own device while recording.
A face detector also runs on the device. If a face appears, the recording stops and is thrown away before anything is uploaded.
On the server every clip is checked: the hand data must show hands in most frames and real movement, and a vision model looks at sampled frames for the right task, hands only, nothing private visible and usable quality. Clips the checks are not sure about are looked at by a person.
Consent
Every contributor confirmed they are 18 or older and agreed to publish their clip and hand data under CC BY 4.0 before uploading.
Contributors get a private code when they first upload. With it they can delete their clips at any time; deleted clips leave the live dataset straight away and are left out of every snapshot built after that.
If you keep a copy of the dataset, please refresh it from a recent snapshot from time to time so deletions reach your copy too.
Format
manifest.json lists every accepted clip with its task, contributor nickname, duration, size and file links.
landmarks/<task>.jsonl has one JSON object per line: the clip id, the task and the full landmark recording (format mimic.landmarks/1).
Daily snapshots are .tar.gz files with the manifest, the landmark files, the clips and this card.
Known limits
Landmarks are estimates from a single camera and can jitter, swap hands or miss fingers that are hidden behind objects.
Contributors are volunteers on their own devices, so lighting, angle and camera quality vary.
Objects in the scene are not labelled; only the task is known.