Skip to content

Tactile data for robot manipulation

Robots can see. They still can't feel.

We record human hands doing real manipulation work — full-hand pressure across 169 taxels, synchronised to egocentric RGB-D and 3D hand pose — and deliver it as training data for manipulation policies.

Tell us what you're training and we'll send sample sessions.

Full-hand pressure · one grasp

colour encodes pressurelowhigh

  1. approach

  2. contact

  3. peak

  4. hold

  5. release

169 TAXELS 16×16 FPC gridCOVERAGE fingers + palmRATE 30 HzMODALITY normal pressure

In every session

Four streams, one clock.

Full-hand pressure
Normal pressure across 169 taxels covering the fingers and palm, sampled through contact, hold and release — not just at the moment of grasp.
Egocentric RGB-D
Colour and per-pixel depth of the same task on a shared clock, so a contact event lines up with what the hand was doing and where the surface it touched actually was.
Hand pose
Per-frame 3D hand tracking, so contact can be located on the object as well as in time.
Task structure
Segment labels for the phases of each task, applied under review rather than inferred after the fact.

Why touch

Vision tells a policy where an object is. It doesn't say whether the hand is touching it, or how hard.

Those are contact questions, and they are the ones that decide whether a grasp succeeds. A policy trained only on pixels has to infer contact from appearance — precisely the inference that fails on deformable objects, on heavy ones, and at the moment the hand itself occludes the thing it is holding.

Manipulation datasets today are still overwhelmingly visual, because touch is harder to instrument. That gap is the opportunity. The same demonstrations, captured with touch, carry signal that no amount of additional camera coverage recovers.

From hand to dataset

Nothing ships that hasn't been reviewed.

  1. Capture

    An operator wears the glove and performs the task. Pressure, video and pose are recorded together, against one clock.

  2. Review

    Every session is inspected before it can be released. Sessions that fail structural or quality checks do not ship.

  3. Release

    Approved sessions are versioned, and a released version is immutable. What you cite is what you get back.

  4. Export

    Delivered as MCAP logs, or as a LeRobot dataset that loads into an existing training pipeline.

Train on the signal you're missing.

We're taking on a small number of partner labs for the first release. Early partners help shape which tasks we capture.