Teleoperation · training data

Teleoperate robot hands. Collect the data to train them.

Grasp is an instrumented glove. Move your fingers and a robot hand moves with them — in simulation today, on a Sharpa Wave and a Unitree Dex5-1. The same glove that drives the robot records the session, at the finger joints, as training data.

In simulation · one session, editedSharpa Wave hand, moved by a human hand in the glove

Working prototype · teleoperation in simulation

NVIDIA Inception Program member
$150
parts, per glove, today
<$80
parts, per pair, projected at 1,000 units
14
calibration gestures, any hand

01 · Teleoperation

Your hand drives a Unitree Dex5-1.

The glove’s joint angles drive Unitree’s five-finger hand in simulation: your hand closes, the robot hand closes. The glove drives the fingers and the hand’s orientation; position comes from the keyboard, because potentiometers measure joints, not where your arm is.

In simulation · nothing sped upPosition from the keyboard — the wrist tracker is the next revision
What it records, and what it misses →

Three robot hands

One glove, three robot hands.

Each robot hand has different proportions, so the glove is calibrated to the wearer and retargeted to the hand. All three clips are simulation, recorded from a person wearing the glove.

  • Sharpa Wave

    Pours coffee · simulated

  • Shadow Hand

    Opens a door · simulated

  • Unitree Dex5-1

    Plays Für Elise · simulated

02 · The data

One take, two views, same frame.

The same glove that drives the robot records the session as training data. Left: the glove on a hand. Right: what the software reconstructed from its channels at that instant, thumb included. Both recordings come from one session, aligned by cross-correlating motion — nothing is re-enacted.

GloveReconstruction
Unedited · single sessionLive stream

What you can buy

A training dataset for your task.

In the paid pilot, we use our glove and capture software to collect demonstrations of your task. Your team gets a licence to use the resulting dataset for robotics training. No outside team has run a pilot yet.

The full price list, glove and camera waitlist included →

Where this goes

Where we want to go.

Video shows where a hand went, but not how a joiner holds a chisel. We want to record skilled people at work, starting with the trades below.

The glove worn on a hand driving a screwdriver into a device: rotary potentiometers at the finger joints and thumb linkage, with the wrist-mounted board behind
The glove in use on a real task
  • Joinery
  • Watchmaking
  • Leatherwork
  • Kitchen work
  • Your domain

We have not recorded any of these at volume yet. That needs a glove cheap enough to give to a hundred people.

Who is building this

  • Łukasz Sojka, Co-founder of Grasp

    Łukasz Sojka

    Co-founder

  • Dmytro Nalyvaiko, Co-founder of Grasp

    Dmytro Nalyvaiko

    Co-founder

A robot hand learns to work from a person showing it how. We built a glove that lets you do that with your own hand — sixteen sensors on the finger joints, about $150 in parts — and today it drives a Unitree Dex5-1 in simulation. The same glove records the session, so the hand that drives the robot also produces the data to train it.
— Łukasz Sojka

Access

Working with robot hands?

Tell us which hand and which task. We will say what works today, in simulation, and what your case would take. If you are a craftsman willing to be recorded, write too.

No newsletter. We only write when there's data to share.