robots learn from hands like yours

Teach a robot how you pour the tea.

Palmline gathers short phone videos of everyday hand tasks, each one recorded to a mission brief. Clips are reviewed, split into labelled steps, and packaged as datasets for teams building robot manipulation.

Rewards for contributors are planned, not live. See the draft.

1 reach 2 grasp 3 lift 4 place
Fig. 1One demonstration, four steps. The path of the hand is what a robot policy learns to imitate.
the whole pipeline, left to right

Brief, Record, Review, Dataset

Four stages. Each one has a person checking it before the next begins.

  1. 1

    Brief

    A mission describes one task: the objects, the start state, the camera angle and how long to take. No guessing what we want.

    read this first
  2. 2

    Record

    Prop your phone, follow the steps, film your hands. A practice page checks framing, light and steadiness before you commit.

    20 to 90 seconds
  3. 3

    Review

    Reviewers check that hands are visible and the task is complete, then mark where each step starts and ends.

    humans in the loop
  4. 4

    Dataset

    Accepted clips are packaged with step labels and metadata, versioned, and delivered to robotics teams.

    versioned releases
two sides of the same notebook

Who it is for

Contributors

Anyone with a phone, a table and ten spare minutes. You do ordinary things, like folding a towel or pouring water, in your own home.

Join the waitlist

Robotics teams

Researchers and engineers who need varied, well-labelled human demonstrations of everyday manipulation to train or evaluate policies.

Request a sample pack

Reviewers

Careful people who watch clips and label steps. The quality of the dataset is the quality of this stage, so it gets the most attention.

See what gets reviewed
what a team receives

A dataset record, in plain view

Every clip ships with a small, readable record. This is Palmline's own draft schema; field names may change before any real release.

  • Step labels with start and end times in milliseconds
  • Camera angle, lighting and framing notes
  • Review outcome and reviewer count
  • Licence and consent flags per clip
{
  "clip_id": "plm_demo_000412",
  "mission": "M-001 pour-water-glass",
  "angle": "overhead-45",
  "duration_ms": 41800,
  "fps": 30,
  "hands_visible_pct": 96.2,
  "steps": [
    { "label": "reach",    "t0": 0,     "t1": 5200  },
    { "label": "grasp",    "t0": 5200,  "t1": 9100  },
    { "label": "pour",     "t0": 9100,  "t1": 31000 },
    { "label": "set_down", "t0": 31000, "t1": 41800 }
  ],
  "review": { "status": "accepted", "reviewers": 2 },
  "consent": { "research_use": true, "faces_blurred": true },
  "split": "train"
}
Fig. 2Draft record. Invented example, not a real clip.
questions in the margin

FAQ

Can I get paid today?

No. Rewards are planned, not live. The rewards page shows a draft of how we expect it to work, but nothing is being paid and nothing is promised.

Does the practice page upload my video?

No. It shows your camera in your own browser tab and analyses frames locally. Nothing is recorded, stored or sent anywhere.

Do my face and home appear in the data?

Missions ask you to frame hands and the work surface only. The draft process includes a review step that rejects clips showing faces or documents.

What kinds of tasks are there?

Kitchen, laundry, desk, tools and care tasks. The missions page lists over thirty invented examples with difficulty, duration and camera angle.

Is the data on this site real?

No. This is a concept demo. Missions, counts and records are simulated.

What is the plotter for?

It is a toy that shows how a two-joint robot arm turns a drawn path into joint angles, the same idea a robot uses to follow a recorded hand path.