From your hands to a robot
Robots that work in homes need to see how people do everyday tasks. Mimic collects short demonstrations from anyone with a phone and gives them back as open data.
Why hands
A robot hand that folds a towel or opens a jar has to know where each finger goes and when. Videos of people show that, but only if the hands are tracked precisely, frame by frame. Mimic saves exactly that: the clip, plus the position of 21 points on each hand for every frame.
What happens when you record
- You pick a task and allow the camera. The camera stays off until you press the button.
- Two small models load into your browser: a hand tracker and a face detector (MediaPipe). They run on your device. The video is not streamed anywhere while you record.
- You see your hands traced live inside a box. After a 3, 2, 1 countdown, Mimic records for 10 seconds.
- If the face detector sees a face, the recording stops straight away and is thrown away. Nothing is uploaded. Point the camera down at the table and try again.
- You review the result: a robot hand replays your movement, and you can watch your clip. Retake as often as you like.
- Only when you tick the boxes and press send does the clip leave your device.
The check
Every clip is checked before it goes public:
- The hand data: hands must be visible in most frames, they must actually move, and the clip must be the right length.
- A vision model looks at a few still frames: is it the right task, are only hands visible, is anything private in view (a screen, a letter, a card), and is it sharp enough.
- A person looks at anything the checks are not sure about.
You get the answer on screen, usually within a minute: accepted, rejected with the reason, or waiting for a person. Rejected clips are deleted from the server.
What gets published
Accepted clips are re-encoded on the server: no sound, no file metadata (like location or device), at most 720p. The clip, its hand data, the task and your nickname go into the open dataset under CC BY 4.0. Anyone can download it from the dataset page.
Your contributor code
With your first upload you get a private code. Keep it. It ties your next uploads to the same nickname, and it lets you delete your clips at any time on the privacy page.
Holders
People who hold $MIMIC can record two extra task packs, get a badge on the leaderboard and vote on the next task each week. See the token page.
Tips for a good clip
- Good light from the front or the side. Avoid strong light behind your hands.
- Prop the phone up so both hands are free, or film from above.
- Clear the table of letters, bills, screens and anything with your name on it.
- Do the task at a normal pace. Slow and clear beats fast.
- 10 second clips of hands doing a task
- 21 tracked points per hand, every frame
- Face in view: recording stops, nothing sent
- Checked by code, a vision model and people
- Free under CC BY 4.0
- Delete your clips any time with your code