Synchronised video, hand pose and fingertip touch from skilled human work in Central Asia, captured with our own hardware and delivered in formats robot-learning teams already train on.
Hardware, software and annotation in one loop, so every recording makes the next one better.
Tactile as ground truth
Contact is measured, not inferred from pixels. Every fingertip reports whether it touches something and how hard.
Force over time separates a grasp from a hold, an adjustment and a release, the details a video-only dataset has to guess.
Contact transitions set the action boundaries, so every segment starts and ends where the hand really changes what it is doing.
Sample recordings come from the capture software running with simulated glove input, shown here to illustrate the data format.
Synchronisation
Vision, touch, hand motion and task labels are stamped on a shared clock, so a pressure spike lands on the exact video frame where the fingers meet the object.
offset target < 5 msData capture system
A thin, stretchable glove with a 4 × 4 pressure array under each fingertip, bend sensing and a motion sensor.
Head-mounted wide-angle camera and two wrist cameras on one hub, so the worker's view and both hands are recorded together.
One app to record, synchronise, review and export, with an on-screen contact view and robot-hand replay.
Multi-device capture
Soon the worker also wears a head unit and two wristbands. Their motion is written to the same clock as the gloves and cameras, so a task is recorded from the eyes to the fingertips.
Gloves are running today. The head unit and wristbands are in development and are not part of any delivered dataset yet.
Session viewer
Drag the timeline. Force, finger bend and the detected phase move together. The empty lanes show where head and wrist motion will appear.
Blue: finger bend. Yellow: fingertip force.
What you receive
| Stream | Source | Status |
|---|---|---|
| Fingertip pressure, 80 pixels per hand | Tactile glove | Running |
| Finger bend and hand joints | Tactile glove | Running |
| Hand orientation and acceleration | Tactile glove | Running |
| Task phases from contact | Studio software | Running |
| Episode quality score | Studio software | Running |
| Egocentric video | Vision rig | Integrating |
| Wrist video, both hands | Vision rig | Integrating |
| Head orientation | Head unit | In development |
| Wrist orientation | Wristbands | In development |
Delivered in LeRobot format today. Other formats on request.
Why touch and body motion
Roadmap
Touch, bend and motion recorded, checked and exported.
Head and wrist cameras on one shared trigger.
Head and wrist motion on the same clock as everything else.
Scoped with a small number of robotics teams.
Replay
Every episode can be played back on a simulated robot hand. Bend angles drive the joints and fingertip force turns the tip red. Shown: a Shadow Hand model in MuJoCo, driven by simulated glove input.
Galafinger Labs turns the touch of skilled human hands into data that robots can learn from. We build the glove, the software and the recording process ourselves.
Get in touch
We are choosing a small number of first partners. Robotics teams: describe the task and the hand you train for. Workshops and factories in Uzbekistan: offer a place to record.
Datasets
Collections are recorded to order. Each one is scoped with you first, then every episode passes quality scoring and a simulated-hand replay before delivery. Nothing below has been recorded yet: pick the tasks you want recorded first.
Questions
A tactile glove (16 pressure pixels per fingertip, bend sensing, hand motion) and a head-mounted wide-angle camera. Wrist cameras are being integrated into the same rig.
Skilled manual work on business premises: tasks that take more than ten minutes to learn, done for real and not staged. No personal data in frame, and written consent from the site owner and every worker.
Models propose hand pose, objects and task phases. Contact transitions set the action boundaries, and a reviewer confirms or corrects each episode. Every episode gets a quality score.
Terms are set in the agreement for each collection. Our standard offer gives the commissioning partner an exclusive period on the tasks it ordered, followed by wider licensing.
LeRobot-format datasets (Parquet with metadata) today. Other formats such as HDF5 or MCAP are available on request, together with a data card for each batch.
Each collection is quoted after scoping, based on the task, the number of hours and the modalities. You receive a sample before you commit.
Galafinger Labs